Research ArticleQuantitative Image Analysis of Epithelial andStromal Area in Histological Sections of Colorectal CancerAn Emerging Diagnostic Tool

R Rogojanu12 T Thalhammer1 U Thiem1 A Heindl1 I Mesteri3 A Seewald4

W Jaumlger5 C Smochina6 I Ellinger1 and G Bises1

1Department of Pathophysiology and Allergy Research Medical University of Vienna 1090 Vienna Austria2TissueGnostics GmbH 1020 Vienna Austria3Clinical Institute of Pathology Medical University of Vienna 1090 Vienna Austria4Seewald Solutions 4616 WeisskirchenTraun Austria5Department of Clinical Pharmacy and Diagnostics University of Vienna 1090 Vienna Austria6Faculty of Automatic Control and Computer Engineering ldquoGheorghe Asachirdquo Technical University of Iasi 700050 Iasi Romania

Correspondence should be addressed to R Rogojanu radurogojanucom

Received 23 July 2015 Revised 28 September 2015 Accepted 30 September 2015

Academic Editor Hung-Ming Lam

Copyright copy 2015 R Rogojanu et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

In colorectal cancer (CRC) an increase in the stromal (S) area with the reduction of the epithelial (E) parts has been suggestedas an indication of tumor progression Therefore an automated image method capable of discriminating E and S areas wouldallow an improved diagnosis Immunofluorescence staining was performed on paraffin-embedded sections from colorectal tumors(16 samples from patients with liver metastasis and 18 without) Noncancerous tumor adjacent mucosa (119899 = 5) and normalmucosa (119899 = 4) were taken as controls Epithelial cells were identified by an anti-keratin 8 (K8) antibody Large tissue areas (5ndash63mm2slide) including tumor center tumor front and adjacent mucosa were scanned using an automated microscopy system(TissueFAXS) With our newly developed algorithms we showed that there is more K8-immunoreactive E in the tumor centerthan in tumor adjacent and normal mucosa Comparing patients with and without metastasis the ES ratio decreased by 20 inthe tumor center and by 40 at tumor front in metastatic samplesThe reduction of Emight be due to a more aggressive phenotypein metastasis patients The novel software allowed a detailed morphometric analysis of cancer tissue compartments as tools forobjective quantitative measurements reduced analysis time and increased reproducibility of the data

1 Introduction

Colorectal cancer (CRC) is a heterogeneous and multifacto-rial disease like other solid tumors [1 2] This might explainwhy tumors identical in the morphology have differentresponses to therapeutic interventions leading to alterationsin the survival rates [3ndash5] For a more precise classificationof the tumors as well as a better estimation of the prognosispotent molecular andor protein markers are clearly neededin addition to the currently used histological grading (G) andTumor-Node-Metastasis (TNM) staging system

More and more the importance of the stroma (S) sur-rounding the epithelial cells (E) is recognized [6] Many

histological studies on S as the tumormicroenvironment havebeen undertaken to investigate the cellular heterogeneity inCRC [7] Thereby the interplay between E and S structuressuch as matrix components and immune cells has beeninvestigated [8] For example tumor growth was found to beinfluenced depending on the subset of tumor infiltrating lym-phocytes [9] and presence of tumor associated macrophageswas associated with an improved overall survival [10] Morewidespread venous and lymphatic invasion is a prognosticparameter for metastases [11] Finally an increased area ofS within the tumor and a decrease in the ES ratio predictspoor survival in CRC [12 13] as well as in breast [14]and in esophageal cancer patients [15] The ES ratio was

Hindawi Publishing CorporationBioMed Research InternationalVolume 2015 Article ID 569071 9 pageshttpdxdoiorg1011552015569071

2 BioMed Research International

therefore suggested to be included in the individual riskestimation [12] So far all data on ES were determinedmanually (semiquantitative) on Haematoxylin and Eosin(HampE) stained sections in the most invasive tumor area by2-3 experts [13ndash15]

The precision of analysis was improved by morphometricmethods for example point counting onHampE stained tumorsections with a virtual Graticule software where objects wereincluded in categories like tumor stroma vessel inflamma-tion and so forth [12]

However the reproducibility of the visual (human-based)quantitative evaluation is time consuming and difficult toachieve while quick qualitative assessment (scoring) may bemore error-prone and with higher interobserver variability

To overcome these problems the aim of this study wasto establish a method for the automated analysis of E and Sin large tumor sections To generate a fast and reproduciblequantification of ES we performed immunofluorescencestaining studies using the keratin 8 marker for E in paraffin-embedded sections fromCRC patients with and without livermetastasis We developed a novel automated microscopicimage analysis method based on segmentation for E andS quantification and used the immunofluorescence (IF)technique to stain the tissue This would allow quantificationof additional markers in relation to epithelial compartmentby a multistaining IF process such as proximity to tumor orratio of expression inside epithelium versus stroma

2 Material and Methods

21 Patients We studied formalin-fixed and paraffin-embedded tumor samples from CRC patients with (119899 = 12)and without (119899 = 14) liver metastasis undergoing surgicalcurative resection at the Department of Surgery at theGeneral Hospital of Vienna between 1995 and 2007 None ofthe patients received preoperative chemo- or radiotherapyAll patients had grade 2 (G2) tumors belonging to a largeheterogeneous group of patients with respect to the othergrades which would benefit from better prognostic andpredictive factors [3] The clinicopathological characteristicsof the patients are shown in Table 1

As controls colon mucosa samples distant from the can-cerous areas were used (119899 = 5) Normal colon tissue sampleswere taken from patients undergoing gastric bypass surgery(119899 = 4) Informed consent was obtained from all patientsand the study was approved by the Ethics Committee of theMedical University of Vienna (protocol number 3582010)

22 Immunofluorescence Staining Following deparaffiniza-tion and rehydration tissue sections (4 120583m) were boiledfor 20 minutes in 005 citraconic anhydride for antigenretrieval

After permeabilization with phosphate buffered saline(PBS)02 Tween and blocking with 5 goat serum(Dianova GmbHJackson ImmunoResearch USA) in 005TweenPBS sections were incubated with rabbit anti-humankeratin 8 antibody (clone EP1628Y Thermo Fisher ScientificFremont CA USA) for 1 hour followed by incubationwith DyLight 549 goat anti-rabbit (Vector Laboratories

Table 1 Clinicopathological characteristics of the study population

minuslivermetastasis

+livermetastasis

Number of patients 14 12Age median (range) 68 (53ndash78) 68 (35ndash79)Male () 12 (86) 8 (67)TNM classificationpTpT1-2 6 0pT3-4 8 12

pNlowast

pN0 13 6pN1-2 1 6

Dukes classificationA 4 0B 8 6C 1 5D 0 1Unknown 1 0LocationCoecum 3 5Colonascendenstransversumdescendens 131 010

Sigmarectum 22 23Unknown 2 1lowastNo patients have more than 2 lymph node metastases

Burlingame CA USA) for 1 hour To verify validity ofK8 as marker for epithelial cells which might acquiremesenchymal properties 120573-catenin double immunolabellingwas performed For double staining after permeabilizationand blocking (as described above) sections were incubatedwith rabbit anti-human 120573-catenin (Abcam Cambridge UK)overnight at 4∘C followed by incubation with DyLight 549goat anti-rabbit IgG (Vector Laboratories Burlingame CAUSA) For the second staining sections were incubatedwith mouse anti-human keratin 8 antibody (Clone TS1Thermo Fischer Scientific Fremont CA USA) for 1 hourand as a secondary antibody Alexa Fluor 647 goat anti-mouse IgG (Invitrogen Molecular ProbesLife TechnologiesPaisley UK) was used 410158406-Diamino-2-phenylindole (DapiRoche Diagnostics GmbH Vienna Austria) was applied at02 120583gmL in PBS for 10min to visualize cell nuclei Thenegative control was prepared by either omitting or replacingthe primary antibodies by nonimmune rabbit andmouse IgGrespectively After washing the slides were mounted withFluoromount G (SouthernBiotech Birmingham AL USA)

All samples were processed under standardized condi-tions to minimize results variations due to the technicalprocedure For automated tissue segmentation by EPSTRAepithelial cells were detected by using only the singleimmunolabelling of keratin 8 (K8)

If not stated otherwise chemicals were obtained fromSigma-Aldrich (St Louis MO USA)

BioMed Research International 3

23 Acquisition with TissueFAXS Tissue sections wereacquired using a TissueFAXS plus tissue cytometer (Tis-sueGnostics GmbH Vienna Austria) using the 20x magni-fication in fluorescence scanning mode For segmentationpurposes (entire tissue epithelial area and lumen) fluo-rescence intensity information from 4 different fluorescencechannels (DAPI FITCCy2 mCherryTxRed and Cy5) wascollected regardless whether a specific staining was presentor not (see also Region-Based Segmentation Approach) Inthe double immunolabelling experiments signals from theCy5 channel (K8) mCherryTxRed (120573-catenin) and DAPI(nuclei) were collected The DAPI staining pattern of nucleiallows the estimation about the quality of the tissue andthe efficacy of the staining procedure in individual samplesThe following filter sets (Chroma 49000 series filter sets)were used DAPI (350 nm excitation 400 nm dichroic and460 nm emission) FITCCy2 (470 nm excitation 495 nmdichroic and 525 nm emission) mCherryTxRed (560 nmexcitation 585 nm dichroic and 630 nm emission) and Cy5(620 nm excitation 660 nm dichroic and 700 nm emission)Previews of the entire slides were taken using a 25x objectiveon the DAPI channel On these images the tissue sec-tions were outlined for scanning with a 20x high powermagnification objective The same autofocus settings wereused for all slides enabling one autofocus point for each3 times 3 group of fields of view To ensure full contrastconditions the integrated extended focus feature was appliedby setting five different z-levels with a 2 120583m interval (2above and 2 below the z-level detected by autofocus) Thez-stack images were merged into one critically sharp imageFluorescence images were acquired and stored with lossycompression in JPEG format 95 quality index For eachField of View (FOV) individual monochrome images werecaptured separately for all selected pixel shift-free filter setsFor presentation false colors replaced the monochromeimage

24 Definition of the Regions of Interest (ROIs) The medianscanned area per slide was 205mm2 (ranging from 64 to315mm2) The area included in individual ROIs (defined bythe user) was 1ndash26mm2 depending on the structure of thesection

On the whole slide overview of the image section ROIswere drawn manually using the zooming and mark-up toolsincluded in the StrataQuest 50 software part of TissueFAXSplus solution The need to analyze and compare differentregions within the tumor tissue was derived from the obser-vation that the proliferation compartment [16] as well as theexpression of proteins like 120573-catenin and cyclin D1 [17] wasnot equally distributed in the tumor mass Heterogeneitiesin the structures were also seen in noncancerous parts of themucosa in CRC samples [18]

To better illustrate the selection of the regions based ontheir characteristic appearance a Hematoxylin and Eosinstained section of a CRC tissue was divided into differentareas (Figure 1)

Under supervision of an experienced pathologist (IMesteri) the areas are categorized as follows

Figure 1 Hematoxylin and Eosin stained section of a CRC tissuewith regions drawn as follows adjacent mucosa region is framed ingreen tumor center in red invasive front 1 in yellow and invasivefront 2 in pink

(i) Normal mucosa from healthy patients (scanned area5ndash9mm2)

(ii) Normal-appearing mucosa (scanned area 4ndash12mm2) distant from the cancerous areas in CRCsamples

(iii) Adjacent mucosa the normal-appearing mucosaadjacent to the tumor (depicted in green in Figures1 2(a) 2(b) and 2(c) scanned area 1ndash26mm2)

(iv) Tumor center central part of the tumors with denseglandular structures surroundedby S locatedgt05 cmaway from the tumor border (depicted in red inFigures 1 2(a) 2(b) and 2(d) scanned area 1ndash24mm2)

(v) Invasive front 1 the invasive margin of the tumorarea where the glandular structures reach out into thesubmucosal S but still form a compact mass (depictedin yellow in Figures 1 2(a) 2(b) and 2(e) scannedarea 1ndash8mm2)

(vi) Invasive front 2 the tumor area facing the submucosalS with infiltrating E structures composed of single K8positive cells or small K8 positive cell clumps (up to10 cells) (depicted in dark pink in Figures 1 2(a) 2(b)and 2(e) scanned area 0ndash2mm2)

The regions were drawn around the epithelial cells locatedon the border of the tumour area to the stroma to avoid anyinclusion of areawhich contains exclusively stroma Artifactssuch as folded or broken tissue areas or areas showing strongbackground staining due to autofluorescence of erythrocytesand lipofuscin were excluded from the analysis

25 NewAlgorithms Themost important and difficult step inautomatic image analysis is the image segmentation whichprovides critical information for further image understand-ing The aim of the new segmentation algorithms is tohighlight the E and S areas as well as the lumen for aproper computation of the ES ratio The outlining shouldbe accurate enough even in tumor areas that have thin Scompartments By analyzing the content of scanned imagesthe following challenges were noticed

(a) Different grey values for the positive areas (nonuni-form staining)

4 BioMed Research International

(a) (b) (c)

(d) (e) (f)

(g) (h)

Figure 2 Example of an image of a CRC tissue section analyzed by EPSTRA (a) K8 staining (TxRed) overlaid with DAPI (blue) includingCy2 (green) and Cy5 (white) background fluorescence (b) epithelial mask (yellow) and stromal mask (green) as found by EPSTRA (c) theadjacent mucosa is depicted in green contours (d) tumor center in red (e) invasive front 1 in yellow and invasive front 2 in pink (f)ndash(h)show the corresponding epithelial mask of (c)ndash(e) views Bubbles show the region of interest (ROI) identification number and its area Barindicates 200120583m

(b) Some cell clusters containing very few cells (they mayrepresent tumor buds)

(c) Holes in the K8 grayscale channel due to lack of K8antigen inside the nuclear compartment

(d) Some nonspecific areas appearing more intenselystained than the positive K8 (erythrocytes auto-fluorescence lipofuscin folded tissue nonuniformtissue reactivity etc)

(e) Multiple intensity-classes of pixels each displayingmultiple peaks in the histogram lumen S weak Eintense E folded S folded E and erythrocytes

(f) Erythrocytes being intense in both TxRed and GFPchannels

(g) E areas which can be very close to each other Inthe center of the tumor the S may only form a verynarrow band between crypts (as thin as 5 pixels wide)

(h) Big data (up to 100000 images per slide each imagewith 14 mega-pixels)

The characteristics and novelty of the proposed method aregiven by the way of integrating the particularity of thesebiological signals (the four input channels and the intensities

BioMed Research International 5

IDAPI

ITexa

IFITC

ICy5

IeDAPI

IeTexa

IeFITC

IeCy5

VCtissue

VCepi

L mask

Tmask

Smask

Emask

AGSm(VCtissue1199831min1199831max)

AGSm(VCepi1199832

min1199832max) cap Tmask

Figure 3 Processing data flow

distribution for E S and lumen) into widely accepted fastimage preprocessing and segmentation methods

26 Region-Based Segmentation Approach Since our goalwas to find the boundaries for the areas of interest we usedregion-based segmentation based on Otsursquos thresholdingmethod [19] Let one consider that 119879ref = thrOtsu(119868) is thethreshold computed with this method on the entire inputimage 119868 Segmenting the entire image with the same globalthreshold (119879ref ) could cause the loss of some epithelial orstroma areas due to intrastaining variability Therefore weused an adaptive threshold applied on image tiles (119879

119894) as

provided by the TissueFAXS file formatA disadvantage of Otsu thresholding is that it separates

the image in two pixel classes even though the image mayhave only one class of pixels (eg stroma could be detectedwithin a tile only with lumen) In addition it may fail whenprovided with images with a multitude of the classes presentin the images (eg lumen weak stroma intense stroma andepithelium)Thus we implemented a restriction such that thethreshold is accepted only if it is within a predefined intervalthrOtsu(119879119894) isin [Tref minus 120572Tref + 120572] where 120572 is a constantWe chose the reference threshold as the one given by Otsuthresholding applied on the entire virtual slide

If the obtained threshold is outside the accepted restrictedinterval the process is repeated by considering only the pix-els within the predefined interval [TminTmax] empiricallyestablished after analyzing multiple image sets All the pixelswith the intensity smaller than Tmin and all the pixels withthe intensity higher thanTmax are eliminated Applying thistransformation before the Otsu thresholding guaranties theproper segmentation of the two classes of interest We namedthis approach ldquoAdaptive Guided Segmentationrdquo

AGS (119868TminTmax) (1)

Since the acquired images present artifacts a simple yetefficient enhancement technique was used after the back-ground is removed by subtracting an appropriate predefinedconstant (119888) the intense areas are further emphasized byraising values of the pixels to a certain power (119901 gt 1)empirically determined for each channel

119868119890 = (119868 minus 119888)119901 (2)

None of the required regions (lumen stroma and epithelialregion) can be accurately extracted from a single chan-nel due to aforementioned intensity variability Therefore

a combination of these channels was used to create morerobust ldquovirtual channelsrdquo (VC) Thus a mix of enhancedDAPI TxRed FITC and Cy5 was made for tissue detection(VCtissue) and one for epithelial detection (VCepi)

VCtissue = 119868119890DAPI + 119868119890GFP + 119868119890TxRed + 119868119890Cy5

VCepi = 119868119890TxRed minus 119868119890GFP minus 119868119890Cy5(3)

The algorithmworkflow needed two branches one for lumen(Lmask) versus tissue (Tmask) masks identification and one forstroma (Smask) versus epithelial (Emask) identification withinthe tissue mask (Tmask) The tissue-lumen detection phaseextracts the tissue and lumenmasks respectively by applyingAGS on VCtissue

LmaskTmask = AGS (VCtissueT1

minT1

max) (4)

Subsequently the epithelial and stromamasks were extractedusing AGS on the VCepi only within tissue mask Tmask

SmaskEmask = AGS (VCepithT2

minT2

max)⋂Tmask (5)

AGS generates irregular contours typical to binarizationmethods Therefore we smoothed the contours by applyinga series of closing and opening morphological operators [20]further referred to as AGSm Since larger smoothing filtersmay generate a better smoothing at the cost of losing detailswe preferred using filters not larger than 5 pixels

The final data flow of the implemented segmentationtechnique is presented in Figure 3

Due to the huge amount of data which needed to beprocessed the proposed algorithmsegmentation techniquewas implemented using a parallel approach making use ofthe independent computing units (cores) available on currentprocessors On a 16-core computer an improvement of 14-fold in the analysis speed was noticed compared to sequentialsingle-core approach reaching an average time of 3min persection

27 Measurements and Statistical Processing Masks of E andS areas were saved as images corresponding to original imagetiles StrataQuest 50 was used to create analysis projectsfrom the original images as well as from the newly createdmask files With the module ldquoTotal Area Measurementrdquo thecalculation was done for the areas of two masks for everysingle marked group of regions normal mucosa mucosa

6 BioMed Research International

(a) (b)

(c) (d)

Figure 4 120573-cateninK8 double staining in a colorectal cancer tissue section (a) Merged image 120573-catenin positive cells are visible in red andK8 positive cells are shown in green Nuclei are shown in light blue (all false colors) The same image is shown with individual fluorescencechannels in grey scale (b) K8 (c) 120573-catenin and (d) nuclear staining with DAPI Arrows with solid lines indicate cells with a translocation(cytoplasmic or nuclear) of 120573-catenin Arrows with dashed lines indicate small groups of cells still positive for K8 but no more expressing120573-catenin

adjacent mucosa tumor center invasive front 1 and invasivefront 2

Batch export functionality extracted the final measure-ments in a single Excel file containing measurements of areaof E and S as well as their ratio for each group of region oneach slide Finally the remaining statistics (averages standarddeviations) were calculated in Excel

3 Results

31 K8 as aMarker for Epithelial Cells in CRC Todifferentiatebetween E and S area in CRC sections we used the epithelialcell marker K8 [21] To verify that the expression of K8 isappropriate to label epithelial cells at various cancer stagesin all patient groups we performed on an additional labelingwith 120573-catenin on 5 test slides which were checked visuallyIn normal epithelial cells 120573-catenin is located on the cellularmembrane at the lateral sides of the cells [22] During cancerprogression 120573-catenin accumulates in the cytoplasm andhence enters the nucleus (see examples in Figures 4(a) and4(c))

However the expression of 120573-catenin is lost in someepithelial cells within the tumor (Figures 4(a) and 4(c)arrows with dashed lines) but a strong expression of K8 ismaintained in all epithelial cells within all areas Thus weconclude that K8 can be considered a reliable marker fornormal as well as for cancerous epithelial cells in the colon

32 Region Definition and Assessment of EPSTRA MaskResults StrataQuest can overlay any combination of chan-nels and masks in various colors and transparencies Figure 2shows examples of such overlays in theway theywere used fordefining and categorizing ROIs EPSTRA mask and contourresults could be also overlaid for visual assessment

33 Comparison of EPSTRA versus Results from HumanExperts and Other Automated Methods Evaluation of themethod was done on a test dataset consisting of 24 FOVsfrom all defined compartments and patient groups Whenassessing the capability of algorithms the ground truthis considered to be results of human experts [23] Tocompensate for interobserver variability we considered theopinion of three different human experts By including themseparately in the performance assessment we can comparethe results from each individual against the performance ofthe proposed EPSTRAmethodThe ground truth (GrTr) wascreated by averaging the manual mark-up (average GrTr)which assigned pixels from the entire test dataset to threemasks lumen S and E After eliminating the lumen maskpixels classified by the method and by the averaged GrTras Ewere considered as ldquotrue positivesrdquo (TP) Also we calculatedldquofalse positiverdquo (FP = pixels marked as A by the algorithm butnot by theGrTr) ldquofalse negativerdquo (FN=pixels notmarked as Eby the algorithmbutmarked by theGrTr) and ldquotrue negativerdquo(TN = pixels not marked as E by neither the algorithm

BioMed Research International 7

Table 2 Specificity precision recall and F1 score from three humanexperts (hExp 1ndash3) alternative automated methods (GT GT + MPLT LT + MP and MH) and EPSTRA

Methodexpert Specificity Precision Recall F1 scorehExp1 097 099 098 098hExp2 098 099 097 098hExp3 098 099 098 099GT 097 099 034 051GT + MP 098 099 030 047LT 093 098 075 085LT + MP 096 099 074 085MH 096 099 080 089EPSTRA 097 099 097 098GT global thresholding using cumulative histogram thresholdingMPmor-phologic processing LT local thresholding MH morphological hierarchy

nor by the GrTr) Precision recallsensitivity specificity andbalanced F1 score [24] were calculated as follows

Specificity = TN(TP + FP)

Precision = TP(TP + FP)

Recall = TP(TP + FN)

F1 = 2 lowast Precision lowast Recall(Precision + Recall)

(6)

As alternative automated workflows we implemented Otsuglobal thresholding (GT) and local thresholding variant byprocessing each image tile independently (LT) Both thesetwo binarization methods were enhanced with additionalmorphological postprocessing (MP) to eliminate the holessituated in the place of the nuclei (methods named GT+ MP and LT + MP resp) We also evaluated the newermethod shown in [25] using morphological hierarchy (MH)although it was designed for and tested on healthy tissue onlyConsidering the results of enumerated alternative methodswe calculated the specificity precision recall and F1 scoreData is summarized in Table 2

34 Evaluation of the ES Ratio in the Defined CompartmentsData on the epithelialstromal ratio generated by EPSTRA issummarized in Figure 5

In healthy colon sections the mean ratio of 115 fornormal mucosa (NM) indicates that the amount of E roughlyequals the amount of S This ratio is similar in the normalappearing mucosa (Mu) distant from tumor area in CRCsections (094 for patients with and 106 for patients withoutliver metastasis) In the mucosa adjacent (AM) to the tumorthe mean ratio is slightly higher (133 for patients with and148 for patients without liver metastasis) An even higheramount of E area was found within the tumor center with anES of 167 for patients with and 201 for patients without livermetastasis At the invasive front 1 (IF1) the ES in patients

minusLM+LM

NM Mu AM TC IF1 IF2000

050

100

150

200

250

300

350

Epith

elia

lstro

mal

ratio

NM Mu AM TC IF1 IF2

115106 148 201 169 012

094 133 167 115 012+LM

minusLM

Figure 5 Epithelialstromal ratio in different tissue compartments(normal mucosa (NM) in healthy patients normal appearingmucosa (Mu) distant from tumor adjacent mucosa (AM) tumorcenter (TC) invasive front 1 (IF1) and invasive front 2 (IF2)) incolorectal cancer patients with and without liver metastasis Valuesare given as means plusmn standard deviation

without liver metastasis was 169 and it was reduced in thatwith livermetastasis (115) As expected at the invasive front 2(IF2) where the tissue is disrupted and only single tumor cellsor small clamps of cells are spread in the stroma the ES ratiois very low (mean ratio of 012 for patients with and withoutliver metastasis resp)

4 Discussion

We established an automatic microscopic imaging method tomeasure the E and S area in paraffin-embedded section fromCRC patients The novel algorithms for the calculation of theES ratio offer significant improvements of the histologicalevaluation of immunofluorescent stained tumor sectionsIn comparison with the HampE staining identification ofepithelial cells is increased due to specific labelling of keratin8 Virtual microscopy overlaying techniques bring additionaladvantages when defining regions of interest based on bothmorphology and intensity of the stained markers

Image processing techniques are used mainly becausethey allow large scale statistical evaluation in addition to clas-sical eye screening As noticed in Table 2 the new EPSTRAmethods achieved better results than existing automaticapproaches almost reaching the performance of humanexperts The ldquoGTrdquo method succeeds in extracting morerelevant threshold values but fails in adapting to existinglocal variations of the sample (ie tumor center shows loweroverall expression levels) On the other side ldquoLTrdquo does notfind proper values in areaswithout lumen or stroma EPSTRAuses information from GT but then continues to adapt tolocal variations succeeding in proper segmentation despiteintrinsic expression variability In addition the fact that itcould find both thin S areas inside tumor centers and smallE structures (which may represent tumor buds) embedded inthe S made it superior when compared to all other evaluated

8 BioMed Research International

automated methods In addition it performed at least asgood as the worse human expert delineating contours in adigital slide with a digital pen at original resolution Sincescoring (quick visual qualitative assessment) is typically fasterand more error-prone than manual drawing of contours ourautomated method would perform better than such visualscoring procedures Morphological postprocessing typicallyhelps to remove very small holes or structures (up to 5pixels) However using MP for holesstructures bigger than5ndash10 pixels dramatically decreases contour accuracy and therecognition rate of thin S andor isolated E cells Theseproblems are also specific to MH [25] and will also appearin other approaches that are based on image downscalingHowever EPSTRA successfully avoided these shortcomingsby working with images at original resolution (pixel size =0323 120583m)

Using EPSTRA on the entire image dataset we foundan increase in E in the central part of the tumor as it wasexpected from the uncontrolled growth of E cells At thetumor margins (IF1) the ES ratio in CRC patients withliver metastasis is 32 lower than in patients without livermetastasis These findings suggest that the lower amount ofE at the tumor margin is related to metastasis and could becaused by epithelial-mesenchymal transition (EMT) of tumorcells A similar reduction of ES ratio was noticed in TCas well Thus patients with the reduced ES ratio may havea shorter progression-free survival time as suggested fromprevious visual assessment studies in colon and breast cancer[12ndash14]

In tumor with increases in S areas E clusters seem tobe less compact The invasive margin shows a rupture ofthe E structure and even budding events Single E cells orsmall E cell clusters are floating in the S This histomor-phological structure [26] at the tumor margin is alreadyproposed as a parameter to be included in predicting CRCprognosis [27] Since budding is difficult to assess visuallythe newly designed EPSTRA method proves to be ade-quate for quantification of this surrogate marker Althoughthe method requires more extensive immunostaining andsophisticated imaging than traditional visual histopathologyit offers important benefits It could provide more addi-tional information than simply counting the tumor budsFurthermore EPSTRA could allowmultiple biomarkers to bequantified selectively over specified cell types regardless oftheir abundance

5 Conclusion

In this study we developed the EPSTRA method for auto-mated compartment basedmorphometric analysis and deter-mination of the ES ratios in CRC sections as a means ofadvanced tissue cytometry We demonstrated that EPSTRAcan be used as a powerful tool for objective quantitativemeasurements of tumor structural development with a scal-able analysis time and reproducibility of data In the futureapplying this technique and elaborating more elaboratedscoring procedures together with additional IF markers mayextend diagnostic options leading to a fast and reliablecharacterization of tumor properties in individual patients

Thereby it may help enabling a better personalized treatmentand individual follow-up in the patientrsquos evaluation promot-ing an improvement of the efficacy of a certain therapeuticregimen [28]

Conflict of Interests

R Rogojanu is a PhD student at Medical University ViennaAustria and is working also for TissueGnostics GmbHVienna Austria as Head of Research amp Development

Authorsrsquo Contribution

Writing was coordinated by T Thalhammer and accom-plished by R Rogojanu C Smochina W Jager and GBises Section staining was optimized and performed by UThiem and G Bises Pathological inspection and clinical datawere provided by I Mesteri Characterization of samples andhighlighting of tumor areas were performed by G BisesAlgorithm development was done by R Rogojanu A Heindland A Seewald Validation of the results was done by AHeindl C Smochina and R Rogojanu Entire project wasdesigned and coordinated by I Ellinger T Thalhammer andG Bises

Acknowledgment

Financial support was given by the FFG Bridge-Project 818094

References

[1] C L Sawyers ldquoThe cancer biomarker problemrdquo Nature vol452 no 7187 pp 548ndash552 2008

[2] S Ogino P Lochhead A T Chan et al ldquoMolecular pathologicalepidemiology of epigenetics emerging integrative science toanalyze environment host and diseaserdquoModern Pathology vol26 no 4 pp 465ndash484 2013

[3] E J A Morris N J Maughan D Forman and P Quirke ldquoWhoto treat with adjuvant therapy in Dukes Bstage II colorectalcancer The need for high quality pathologyrdquo Gut vol 56 no10 pp 1419ndash1425 2007

[4] J R Jass ldquoMolecular heterogeneity of colorectal cancer impli-cations for cancer controlrdquo Surgical Oncology vol 16 supple-ment 1 pp S7ndashS9 2007

[5] I Zlobec and A Lugli ldquoPrognostic and predictive factors incolorectal cancerrdquo Postgraduate Medical Journal vol 84 no994 pp 403ndash411 2008

[6] M Allen and J Louise Jones ldquoJekyll and Hyde the role ofthe microenvironment on the progression of cancerrdquo Journal ofPathology vol 223 no 2 pp 162ndash176 2011

[7] E Karamitopoulou I Zlobec I Panayiotides et al ldquoSystematicanalysis of proteins from different signaling pathways in thetumor center and the invasive front of colorectal cancerrdquoHuman Pathology vol 42 no 12 pp 1888ndash1896 2011

[8] M Sund and R Kalluri ldquoTumor stroma derived biomarkers incancerrdquo Cancer and Metastasis Reviews vol 28 no 1-2 pp 177ndash183 2009

BioMed Research International 9

[9] M Scurr A Gallimore and A Godkin ldquoT cell subsets andcolorectal cancer discerning the good from the badrdquo CellularImmunology vol 279 no 1 pp 21ndash24 2012

[10] Q-W Zhang L Liu C-Y Gong et al ldquoPrognostic significanceof tumor-associated macrophages in solid tumor a meta-analysis of the literaturerdquo PLoS ONE vol 7 no 12 Article IDe50946 2012

[11] J Betge M J Pollheimer R A Lindtner et al ldquoIntramural andextramural vascular invasion in colorectal cancer prognosticsignificance and quality of pathology reportingrdquo Cancer vol118 no 3 pp 628ndash638 2012

[12] N P West M Dattani P McShane et al ldquoThe proportionof tumour cells is an independent predictor for survival incolorectal cancer patientsrdquo British Journal of Cancer vol 102no 10 pp 1519ndash1523 2010

[13] A Huijbers R A E M Tollenaar G W V Pelt et al ldquoTheproportion of tumor-stroma as a strong prognosticator for stageII and III colon cancer patients validation in the victor trialrdquoAnnals of Oncology vol 24 no 1 Article ID mds246 pp 179ndash185 2013

[14] E M de Kruijf J G H van Nes C J H van de Velde et alldquoTumor-stroma ratio in the primary tumor is a prognostic fac-tor in early breast cancer patients especially in triple-negativecarcinoma patientsrdquo Breast Cancer Research and Treatment vol125 no 3 pp 687ndash696 2011

[15] E F W Courrech Staal V T H B M Smit M-L F VanVelthuysen et al ldquoReproducibility and validation of tumourstroma ratio scoring on oesophageal adenocarcinoma biopsiesrdquoEuropean Journal of Cancer vol 47 no 3 pp 375ndash382 2011

[16] C A Rubio ldquoCell proliferation at the leading invasive frontof colonic carcinomas Preliminary observationsrdquo AnticancerResearch vol 26 no 3 pp 2275ndash2278 2006

[17] A Jung M Schrauder U Oswald et al ldquoThe invasion frontof human colorectal adenocarcinomas shows co-localization ofnuclear 120573-catenin cyclin D1 and p16INK4A and is a region oflow proliferationrdquo The American Journal of Pathology vol 159no 5 pp 1613ndash1617 2001

[18] M Ponz de Leon L Roncucci P Di Donato et al ldquoPatternof epithelial cell proliferation in colorectal mucosa of normalsubjects and of patients with adenomatous polyps or cancer ofthe large bowelrdquo Cancer Research vol 48 no 14 pp 4121ndash41261988

[19] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[20] R C Gonzales and R E Woods Digital Image ProcessingPrentice-Hall 3rd edition 2007

[21] V Karantza ldquoKeratins in health and cancer more than mereepithelial cell markersrdquo Oncogene vol 30 no 2 pp 127ndash1382011

[22] T Brabletz F Hlubek S Spaderna et al ldquoInvasion and metas-tasis in colorectal cancer epithelial-mesenchymal transitionmesenchymal-epithelial transition stemcells and beta-cateninrdquoCells Tissues Organs vol 179 no 1-2 pp 56ndash65 2005

[23] S B Wilson and R Emerson ldquoSpike detection a review andcomparison of algorithmsrdquo Clinical Neurophysiology vol 113no 12 pp 1873ndash1881 2002

[24] C J van Rijsbergen Information Retrieval Butterworths Lon-don UK 2nd edition 1979

[25] C Smochina V Manta and W Kropatsch ldquoCrypts detectionin microscopic images using hierarchical structuresrdquo PatternRecognition Letters vol 34 no 8 pp 934ndash941 2013

[26] J P Sleeman G Christofori R Fodde et al ldquoConcepts ofmetastasis in flux the stromal progression modelrdquo Seminars inCancer Biology vol 22 no 3 pp 174ndash186 2012

[27] A Lugli E Karamitopoulou and I Zlobec ldquoTumour buddinga promising parameter in colorectal cancerrdquo British Journal ofCancer vol 106 no 11 pp 1713ndash1717 2012

[28] M R Trusheim E R Berndt and F L Douglas ldquoStratifiedmedicine strategic and economic implications of combiningdrugs and clinical biomarkersrdquoNature Reviews Drug Discoveryvol 6 no 4 pp 287ndash293 2007

BioMed Research International 3

23 Acquisition with TissueFAXS Tissue sections wereacquired using a TissueFAXS plus tissue cytometer (Tis-sueGnostics GmbH Vienna Austria) using the 20x magni-fication in fluorescence scanning mode For segmentationpurposes (entire tissue epithelial area and lumen) fluo-rescence intensity information from 4 different fluorescencechannels (DAPI FITCCy2 mCherryTxRed and Cy5) wascollected regardless whether a specific staining was presentor not (see also Region-Based Segmentation Approach) Inthe double immunolabelling experiments signals from theCy5 channel (K8) mCherryTxRed (120573-catenin) and DAPI(nuclei) were collected The DAPI staining pattern of nucleiallows the estimation about the quality of the tissue andthe efficacy of the staining procedure in individual samplesThe following filter sets (Chroma 49000 series filter sets)were used DAPI (350 nm excitation 400 nm dichroic and460 nm emission) FITCCy2 (470 nm excitation 495 nmdichroic and 525 nm emission) mCherryTxRed (560 nmexcitation 585 nm dichroic and 630 nm emission) and Cy5(620 nm excitation 660 nm dichroic and 700 nm emission)Previews of the entire slides were taken using a 25x objectiveon the DAPI channel On these images the tissue sec-tions were outlined for scanning with a 20x high powermagnification objective The same autofocus settings wereused for all slides enabling one autofocus point for each3 times 3 group of fields of view To ensure full contrastconditions the integrated extended focus feature was appliedby setting five different z-levels with a 2 120583m interval (2above and 2 below the z-level detected by autofocus) Thez-stack images were merged into one critically sharp imageFluorescence images were acquired and stored with lossycompression in JPEG format 95 quality index For eachField of View (FOV) individual monochrome images werecaptured separately for all selected pixel shift-free filter setsFor presentation false colors replaced the monochromeimage

24 Definition of the Regions of Interest (ROIs) The medianscanned area per slide was 205mm2 (ranging from 64 to315mm2) The area included in individual ROIs (defined bythe user) was 1ndash26mm2 depending on the structure of thesection

On the whole slide overview of the image section ROIswere drawn manually using the zooming and mark-up toolsincluded in the StrataQuest 50 software part of TissueFAXSplus solution The need to analyze and compare differentregions within the tumor tissue was derived from the obser-vation that the proliferation compartment [16] as well as theexpression of proteins like 120573-catenin and cyclin D1 [17] wasnot equally distributed in the tumor mass Heterogeneitiesin the structures were also seen in noncancerous parts of themucosa in CRC samples [18]

To better illustrate the selection of the regions based ontheir characteristic appearance a Hematoxylin and Eosinstained section of a CRC tissue was divided into differentareas (Figure 1)

Under supervision of an experienced pathologist (IMesteri) the areas are categorized as follows

Figure 1 Hematoxylin and Eosin stained section of a CRC tissuewith regions drawn as follows adjacent mucosa region is framed ingreen tumor center in red invasive front 1 in yellow and invasivefront 2 in pink

(i) Normal mucosa from healthy patients (scanned area5ndash9mm2)

(ii) Normal-appearing mucosa (scanned area 4ndash12mm2) distant from the cancerous areas in CRCsamples

(iii) Adjacent mucosa the normal-appearing mucosaadjacent to the tumor (depicted in green in Figures1 2(a) 2(b) and 2(c) scanned area 1ndash26mm2)

(iv) Tumor center central part of the tumors with denseglandular structures surroundedby S locatedgt05 cmaway from the tumor border (depicted in red inFigures 1 2(a) 2(b) and 2(d) scanned area 1ndash24mm2)

(v) Invasive front 1 the invasive margin of the tumorarea where the glandular structures reach out into thesubmucosal S but still form a compact mass (depictedin yellow in Figures 1 2(a) 2(b) and 2(e) scannedarea 1ndash8mm2)

(vi) Invasive front 2 the tumor area facing the submucosalS with infiltrating E structures composed of single K8positive cells or small K8 positive cell clumps (up to10 cells) (depicted in dark pink in Figures 1 2(a) 2(b)and 2(e) scanned area 0ndash2mm2)

The regions were drawn around the epithelial cells locatedon the border of the tumour area to the stroma to avoid anyinclusion of areawhich contains exclusively stroma Artifactssuch as folded or broken tissue areas or areas showing strongbackground staining due to autofluorescence of erythrocytesand lipofuscin were excluded from the analysis

25 NewAlgorithms Themost important and difficult step inautomatic image analysis is the image segmentation whichprovides critical information for further image understand-ing The aim of the new segmentation algorithms is tohighlight the E and S areas as well as the lumen for aproper computation of the ES ratio The outlining shouldbe accurate enough even in tumor areas that have thin Scompartments By analyzing the content of scanned imagesthe following challenges were noticed

(a) Different grey values for the positive areas (nonuni-form staining)

4 BioMed Research International

(a) (b) (c)

(d) (e) (f)

(g) (h)

Figure 2 Example of an image of a CRC tissue section analyzed by EPSTRA (a) K8 staining (TxRed) overlaid with DAPI (blue) includingCy2 (green) and Cy5 (white) background fluorescence (b) epithelial mask (yellow) and stromal mask (green) as found by EPSTRA (c) theadjacent mucosa is depicted in green contours (d) tumor center in red (e) invasive front 1 in yellow and invasive front 2 in pink (f)ndash(h)show the corresponding epithelial mask of (c)ndash(e) views Bubbles show the region of interest (ROI) identification number and its area Barindicates 200120583m

(b) Some cell clusters containing very few cells (they mayrepresent tumor buds)

(c) Holes in the K8 grayscale channel due to lack of K8antigen inside the nuclear compartment

(d) Some nonspecific areas appearing more intenselystained than the positive K8 (erythrocytes auto-fluorescence lipofuscin folded tissue nonuniformtissue reactivity etc)

(e) Multiple intensity-classes of pixels each displayingmultiple peaks in the histogram lumen S weak Eintense E folded S folded E and erythrocytes

(f) Erythrocytes being intense in both TxRed and GFPchannels

(g) E areas which can be very close to each other Inthe center of the tumor the S may only form a verynarrow band between crypts (as thin as 5 pixels wide)

(h) Big data (up to 100000 images per slide each imagewith 14 mega-pixels)

The characteristics and novelty of the proposed method aregiven by the way of integrating the particularity of thesebiological signals (the four input channels and the intensities

BioMed Research International 5

IDAPI

ITexa

IFITC

ICy5

IeDAPI

IeTexa

IeFITC

IeCy5

VCtissue

VCepi

L mask

Tmask

Smask

Emask

AGSm(VCtissue1199831min1199831max)

AGSm(VCepi1199832

min1199832max) cap Tmask

Figure 3 Processing data flow

distribution for E S and lumen) into widely accepted fastimage preprocessing and segmentation methods

26 Region-Based Segmentation Approach Since our goalwas to find the boundaries for the areas of interest we usedregion-based segmentation based on Otsursquos thresholdingmethod [19] Let one consider that 119879ref = thrOtsu(119868) is thethreshold computed with this method on the entire inputimage 119868 Segmenting the entire image with the same globalthreshold (119879ref ) could cause the loss of some epithelial orstroma areas due to intrastaining variability Therefore weused an adaptive threshold applied on image tiles (119879

119894) as

provided by the TissueFAXS file formatA disadvantage of Otsu thresholding is that it separates

the image in two pixel classes even though the image mayhave only one class of pixels (eg stroma could be detectedwithin a tile only with lumen) In addition it may fail whenprovided with images with a multitude of the classes presentin the images (eg lumen weak stroma intense stroma andepithelium)Thus we implemented a restriction such that thethreshold is accepted only if it is within a predefined intervalthrOtsu(119879119894) isin [Tref minus 120572Tref + 120572] where 120572 is a constantWe chose the reference threshold as the one given by Otsuthresholding applied on the entire virtual slide

If the obtained threshold is outside the accepted restrictedinterval the process is repeated by considering only the pix-els within the predefined interval [TminTmax] empiricallyestablished after analyzing multiple image sets All the pixelswith the intensity smaller than Tmin and all the pixels withthe intensity higher thanTmax are eliminated Applying thistransformation before the Otsu thresholding guaranties theproper segmentation of the two classes of interest We namedthis approach ldquoAdaptive Guided Segmentationrdquo

AGS (119868TminTmax) (1)

Since the acquired images present artifacts a simple yetefficient enhancement technique was used after the back-ground is removed by subtracting an appropriate predefinedconstant (119888) the intense areas are further emphasized byraising values of the pixels to a certain power (119901 gt 1)empirically determined for each channel

119868119890 = (119868 minus 119888)119901 (2)

None of the required regions (lumen stroma and epithelialregion) can be accurately extracted from a single chan-nel due to aforementioned intensity variability Therefore

a combination of these channels was used to create morerobust ldquovirtual channelsrdquo (VC) Thus a mix of enhancedDAPI TxRed FITC and Cy5 was made for tissue detection(VCtissue) and one for epithelial detection (VCepi)

VCtissue = 119868119890DAPI + 119868119890GFP + 119868119890TxRed + 119868119890Cy5

VCepi = 119868119890TxRed minus 119868119890GFP minus 119868119890Cy5(3)

The algorithmworkflow needed two branches one for lumen(Lmask) versus tissue (Tmask) masks identification and one forstroma (Smask) versus epithelial (Emask) identification withinthe tissue mask (Tmask) The tissue-lumen detection phaseextracts the tissue and lumenmasks respectively by applyingAGS on VCtissue

LmaskTmask = AGS (VCtissueT1

minT1

max) (4)

Subsequently the epithelial and stromamasks were extractedusing AGS on the VCepi only within tissue mask Tmask

SmaskEmask = AGS (VCepithT2

minT2

max)⋂Tmask (5)

AGS generates irregular contours typical to binarizationmethods Therefore we smoothed the contours by applyinga series of closing and opening morphological operators [20]further referred to as AGSm Since larger smoothing filtersmay generate a better smoothing at the cost of losing detailswe preferred using filters not larger than 5 pixels

The final data flow of the implemented segmentationtechnique is presented in Figure 3

Due to the huge amount of data which needed to beprocessed the proposed algorithmsegmentation techniquewas implemented using a parallel approach making use ofthe independent computing units (cores) available on currentprocessors On a 16-core computer an improvement of 14-fold in the analysis speed was noticed compared to sequentialsingle-core approach reaching an average time of 3min persection

27 Measurements and Statistical Processing Masks of E andS areas were saved as images corresponding to original imagetiles StrataQuest 50 was used to create analysis projectsfrom the original images as well as from the newly createdmask files With the module ldquoTotal Area Measurementrdquo thecalculation was done for the areas of two masks for everysingle marked group of regions normal mucosa mucosa

6 BioMed Research International

(a) (b)

(c) (d)

Figure 4 120573-cateninK8 double staining in a colorectal cancer tissue section (a) Merged image 120573-catenin positive cells are visible in red andK8 positive cells are shown in green Nuclei are shown in light blue (all false colors) The same image is shown with individual fluorescencechannels in grey scale (b) K8 (c) 120573-catenin and (d) nuclear staining with DAPI Arrows with solid lines indicate cells with a translocation(cytoplasmic or nuclear) of 120573-catenin Arrows with dashed lines indicate small groups of cells still positive for K8 but no more expressing120573-catenin

adjacent mucosa tumor center invasive front 1 and invasivefront 2

Batch export functionality extracted the final measure-ments in a single Excel file containing measurements of areaof E and S as well as their ratio for each group of region oneach slide Finally the remaining statistics (averages standarddeviations) were calculated in Excel

3 Results

31 K8 as aMarker for Epithelial Cells in CRC Todifferentiatebetween E and S area in CRC sections we used the epithelialcell marker K8 [21] To verify that the expression of K8 isappropriate to label epithelial cells at various cancer stagesin all patient groups we performed on an additional labelingwith 120573-catenin on 5 test slides which were checked visuallyIn normal epithelial cells 120573-catenin is located on the cellularmembrane at the lateral sides of the cells [22] During cancerprogression 120573-catenin accumulates in the cytoplasm andhence enters the nucleus (see examples in Figures 4(a) and4(c))

However the expression of 120573-catenin is lost in someepithelial cells within the tumor (Figures 4(a) and 4(c)arrows with dashed lines) but a strong expression of K8 ismaintained in all epithelial cells within all areas Thus weconclude that K8 can be considered a reliable marker fornormal as well as for cancerous epithelial cells in the colon

32 Region Definition and Assessment of EPSTRA MaskResults StrataQuest can overlay any combination of chan-nels and masks in various colors and transparencies Figure 2shows examples of such overlays in theway theywere used fordefining and categorizing ROIs EPSTRA mask and contourresults could be also overlaid for visual assessment

33 Comparison of EPSTRA versus Results from HumanExperts and Other Automated Methods Evaluation of themethod was done on a test dataset consisting of 24 FOVsfrom all defined compartments and patient groups Whenassessing the capability of algorithms the ground truthis considered to be results of human experts [23] Tocompensate for interobserver variability we considered theopinion of three different human experts By including themseparately in the performance assessment we can comparethe results from each individual against the performance ofthe proposed EPSTRAmethodThe ground truth (GrTr) wascreated by averaging the manual mark-up (average GrTr)which assigned pixels from the entire test dataset to threemasks lumen S and E After eliminating the lumen maskpixels classified by the method and by the averaged GrTras Ewere considered as ldquotrue positivesrdquo (TP) Also we calculatedldquofalse positiverdquo (FP = pixels marked as A by the algorithm butnot by theGrTr) ldquofalse negativerdquo (FN=pixels notmarked as Eby the algorithmbutmarked by theGrTr) and ldquotrue negativerdquo(TN = pixels not marked as E by neither the algorithm

BioMed Research International 7

Table 2 Specificity precision recall and F1 score from three humanexperts (hExp 1ndash3) alternative automated methods (GT GT + MPLT LT + MP and MH) and EPSTRA

Methodexpert Specificity Precision Recall F1 scorehExp1 097 099 098 098hExp2 098 099 097 098hExp3 098 099 098 099GT 097 099 034 051GT + MP 098 099 030 047LT 093 098 075 085LT + MP 096 099 074 085MH 096 099 080 089EPSTRA 097 099 097 098GT global thresholding using cumulative histogram thresholdingMPmor-phologic processing LT local thresholding MH morphological hierarchy

nor by the GrTr) Precision recallsensitivity specificity andbalanced F1 score [24] were calculated as follows

Specificity = TN(TP + FP)

Precision = TP(TP + FP)

Recall = TP(TP + FN)

F1 = 2 lowast Precision lowast Recall(Precision + Recall)

(6)

As alternative automated workflows we implemented Otsuglobal thresholding (GT) and local thresholding variant byprocessing each image tile independently (LT) Both thesetwo binarization methods were enhanced with additionalmorphological postprocessing (MP) to eliminate the holessituated in the place of the nuclei (methods named GT+ MP and LT + MP resp) We also evaluated the newermethod shown in [25] using morphological hierarchy (MH)although it was designed for and tested on healthy tissue onlyConsidering the results of enumerated alternative methodswe calculated the specificity precision recall and F1 scoreData is summarized in Table 2

34 Evaluation of the ES Ratio in the Defined CompartmentsData on the epithelialstromal ratio generated by EPSTRA issummarized in Figure 5

In healthy colon sections the mean ratio of 115 fornormal mucosa (NM) indicates that the amount of E roughlyequals the amount of S This ratio is similar in the normalappearing mucosa (Mu) distant from tumor area in CRCsections (094 for patients with and 106 for patients withoutliver metastasis) In the mucosa adjacent (AM) to the tumorthe mean ratio is slightly higher (133 for patients with and148 for patients without liver metastasis) An even higheramount of E area was found within the tumor center with anES of 167 for patients with and 201 for patients without livermetastasis At the invasive front 1 (IF1) the ES in patients

minusLM+LM

NM Mu AM TC IF1 IF2000

050

100

150

200

250

300

350

Epith

elia

lstro

mal

ratio

NM Mu AM TC IF1 IF2

115106 148 201 169 012

094 133 167 115 012+LM

minusLM

Figure 5 Epithelialstromal ratio in different tissue compartments(normal mucosa (NM) in healthy patients normal appearingmucosa (Mu) distant from tumor adjacent mucosa (AM) tumorcenter (TC) invasive front 1 (IF1) and invasive front 2 (IF2)) incolorectal cancer patients with and without liver metastasis Valuesare given as means plusmn standard deviation

without liver metastasis was 169 and it was reduced in thatwith livermetastasis (115) As expected at the invasive front 2(IF2) where the tissue is disrupted and only single tumor cellsor small clamps of cells are spread in the stroma the ES ratiois very low (mean ratio of 012 for patients with and withoutliver metastasis resp)

4 Discussion

We established an automatic microscopic imaging method tomeasure the E and S area in paraffin-embedded section fromCRC patients The novel algorithms for the calculation of theES ratio offer significant improvements of the histologicalevaluation of immunofluorescent stained tumor sectionsIn comparison with the HampE staining identification ofepithelial cells is increased due to specific labelling of keratin8 Virtual microscopy overlaying techniques bring additionaladvantages when defining regions of interest based on bothmorphology and intensity of the stained markers

Image processing techniques are used mainly becausethey allow large scale statistical evaluation in addition to clas-sical eye screening As noticed in Table 2 the new EPSTRAmethods achieved better results than existing automaticapproaches almost reaching the performance of humanexperts The ldquoGTrdquo method succeeds in extracting morerelevant threshold values but fails in adapting to existinglocal variations of the sample (ie tumor center shows loweroverall expression levels) On the other side ldquoLTrdquo does notfind proper values in areaswithout lumen or stroma EPSTRAuses information from GT but then continues to adapt tolocal variations succeeding in proper segmentation despiteintrinsic expression variability In addition the fact that itcould find both thin S areas inside tumor centers and smallE structures (which may represent tumor buds) embedded inthe S made it superior when compared to all other evaluated

8 BioMed Research International

automated methods In addition it performed at least asgood as the worse human expert delineating contours in adigital slide with a digital pen at original resolution Sincescoring (quick visual qualitative assessment) is typically fasterand more error-prone than manual drawing of contours ourautomated method would perform better than such visualscoring procedures Morphological postprocessing typicallyhelps to remove very small holes or structures (up to 5pixels) However using MP for holesstructures bigger than5ndash10 pixels dramatically decreases contour accuracy and therecognition rate of thin S andor isolated E cells Theseproblems are also specific to MH [25] and will also appearin other approaches that are based on image downscalingHowever EPSTRA successfully avoided these shortcomingsby working with images at original resolution (pixel size =0323 120583m)

Using EPSTRA on the entire image dataset we foundan increase in E in the central part of the tumor as it wasexpected from the uncontrolled growth of E cells At thetumor margins (IF1) the ES ratio in CRC patients withliver metastasis is 32 lower than in patients without livermetastasis These findings suggest that the lower amount ofE at the tumor margin is related to metastasis and could becaused by epithelial-mesenchymal transition (EMT) of tumorcells A similar reduction of ES ratio was noticed in TCas well Thus patients with the reduced ES ratio may havea shorter progression-free survival time as suggested fromprevious visual assessment studies in colon and breast cancer[12ndash14]

In tumor with increases in S areas E clusters seem tobe less compact The invasive margin shows a rupture ofthe E structure and even budding events Single E cells orsmall E cell clusters are floating in the S This histomor-phological structure [26] at the tumor margin is alreadyproposed as a parameter to be included in predicting CRCprognosis [27] Since budding is difficult to assess visuallythe newly designed EPSTRA method proves to be ade-quate for quantification of this surrogate marker Althoughthe method requires more extensive immunostaining andsophisticated imaging than traditional visual histopathologyit offers important benefits It could provide more addi-tional information than simply counting the tumor budsFurthermore EPSTRA could allowmultiple biomarkers to bequantified selectively over specified cell types regardless oftheir abundance

5 Conclusion

In this study we developed the EPSTRA method for auto-mated compartment basedmorphometric analysis and deter-mination of the ES ratios in CRC sections as a means ofadvanced tissue cytometry We demonstrated that EPSTRAcan be used as a powerful tool for objective quantitativemeasurements of tumor structural development with a scal-able analysis time and reproducibility of data In the futureapplying this technique and elaborating more elaboratedscoring procedures together with additional IF markers mayextend diagnostic options leading to a fast and reliablecharacterization of tumor properties in individual patients

Thereby it may help enabling a better personalized treatmentand individual follow-up in the patientrsquos evaluation promot-ing an improvement of the efficacy of a certain therapeuticregimen [28]

Conflict of Interests

R Rogojanu is a PhD student at Medical University ViennaAustria and is working also for TissueGnostics GmbHVienna Austria as Head of Research amp Development

Authorsrsquo Contribution

Writing was coordinated by T Thalhammer and accom-plished by R Rogojanu C Smochina W Jager and GBises Section staining was optimized and performed by UThiem and G Bises Pathological inspection and clinical datawere provided by I Mesteri Characterization of samples andhighlighting of tumor areas were performed by G BisesAlgorithm development was done by R Rogojanu A Heindland A Seewald Validation of the results was done by AHeindl C Smochina and R Rogojanu Entire project wasdesigned and coordinated by I Ellinger T Thalhammer andG Bises

Acknowledgment

Financial support was given by the FFG Bridge-Project 818094

References

[1] C L Sawyers ldquoThe cancer biomarker problemrdquo Nature vol452 no 7187 pp 548ndash552 2008

[2] S Ogino P Lochhead A T Chan et al ldquoMolecular pathologicalepidemiology of epigenetics emerging integrative science toanalyze environment host and diseaserdquoModern Pathology vol26 no 4 pp 465ndash484 2013

[3] E J A Morris N J Maughan D Forman and P Quirke ldquoWhoto treat with adjuvant therapy in Dukes Bstage II colorectalcancer The need for high quality pathologyrdquo Gut vol 56 no10 pp 1419ndash1425 2007

[4] J R Jass ldquoMolecular heterogeneity of colorectal cancer impli-cations for cancer controlrdquo Surgical Oncology vol 16 supple-ment 1 pp S7ndashS9 2007

[5] I Zlobec and A Lugli ldquoPrognostic and predictive factors incolorectal cancerrdquo Postgraduate Medical Journal vol 84 no994 pp 403ndash411 2008

[6] M Allen and J Louise Jones ldquoJekyll and Hyde the role ofthe microenvironment on the progression of cancerrdquo Journal ofPathology vol 223 no 2 pp 162ndash176 2011

[7] E Karamitopoulou I Zlobec I Panayiotides et al ldquoSystematicanalysis of proteins from different signaling pathways in thetumor center and the invasive front of colorectal cancerrdquoHuman Pathology vol 42 no 12 pp 1888ndash1896 2011

[8] M Sund and R Kalluri ldquoTumor stroma derived biomarkers incancerrdquo Cancer and Metastasis Reviews vol 28 no 1-2 pp 177ndash183 2009

BioMed Research International 9

[9] M Scurr A Gallimore and A Godkin ldquoT cell subsets andcolorectal cancer discerning the good from the badrdquo CellularImmunology vol 279 no 1 pp 21ndash24 2012

[10] Q-W Zhang L Liu C-Y Gong et al ldquoPrognostic significanceof tumor-associated macrophages in solid tumor a meta-analysis of the literaturerdquo PLoS ONE vol 7 no 12 Article IDe50946 2012

[11] J Betge M J Pollheimer R A Lindtner et al ldquoIntramural andextramural vascular invasion in colorectal cancer prognosticsignificance and quality of pathology reportingrdquo Cancer vol118 no 3 pp 628ndash638 2012

[12] N P West M Dattani P McShane et al ldquoThe proportionof tumour cells is an independent predictor for survival incolorectal cancer patientsrdquo British Journal of Cancer vol 102no 10 pp 1519ndash1523 2010

[13] A Huijbers R A E M Tollenaar G W V Pelt et al ldquoTheproportion of tumor-stroma as a strong prognosticator for stageII and III colon cancer patients validation in the victor trialrdquoAnnals of Oncology vol 24 no 1 Article ID mds246 pp 179ndash185 2013

[14] E M de Kruijf J G H van Nes C J H van de Velde et alldquoTumor-stroma ratio in the primary tumor is a prognostic fac-tor in early breast cancer patients especially in triple-negativecarcinoma patientsrdquo Breast Cancer Research and Treatment vol125 no 3 pp 687ndash696 2011

[15] E F W Courrech Staal V T H B M Smit M-L F VanVelthuysen et al ldquoReproducibility and validation of tumourstroma ratio scoring on oesophageal adenocarcinoma biopsiesrdquoEuropean Journal of Cancer vol 47 no 3 pp 375ndash382 2011

[16] C A Rubio ldquoCell proliferation at the leading invasive frontof colonic carcinomas Preliminary observationsrdquo AnticancerResearch vol 26 no 3 pp 2275ndash2278 2006

[17] A Jung M Schrauder U Oswald et al ldquoThe invasion frontof human colorectal adenocarcinomas shows co-localization ofnuclear 120573-catenin cyclin D1 and p16INK4A and is a region oflow proliferationrdquo The American Journal of Pathology vol 159no 5 pp 1613ndash1617 2001

[18] M Ponz de Leon L Roncucci P Di Donato et al ldquoPatternof epithelial cell proliferation in colorectal mucosa of normalsubjects and of patients with adenomatous polyps or cancer ofthe large bowelrdquo Cancer Research vol 48 no 14 pp 4121ndash41261988

[19] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[20] R C Gonzales and R E Woods Digital Image ProcessingPrentice-Hall 3rd edition 2007

[21] V Karantza ldquoKeratins in health and cancer more than mereepithelial cell markersrdquo Oncogene vol 30 no 2 pp 127ndash1382011

[22] T Brabletz F Hlubek S Spaderna et al ldquoInvasion and metas-tasis in colorectal cancer epithelial-mesenchymal transitionmesenchymal-epithelial transition stemcells and beta-cateninrdquoCells Tissues Organs vol 179 no 1-2 pp 56ndash65 2005

[23] S B Wilson and R Emerson ldquoSpike detection a review andcomparison of algorithmsrdquo Clinical Neurophysiology vol 113no 12 pp 1873ndash1881 2002

[24] C J van Rijsbergen Information Retrieval Butterworths Lon-don UK 2nd edition 1979

[25] C Smochina V Manta and W Kropatsch ldquoCrypts detectionin microscopic images using hierarchical structuresrdquo PatternRecognition Letters vol 34 no 8 pp 934ndash941 2013

[26] J P Sleeman G Christofori R Fodde et al ldquoConcepts ofmetastasis in flux the stromal progression modelrdquo Seminars inCancer Biology vol 22 no 3 pp 174ndash186 2012

[27] A Lugli E Karamitopoulou and I Zlobec ldquoTumour buddinga promising parameter in colorectal cancerrdquo British Journal ofCancer vol 106 no 11 pp 1713ndash1717 2012

[28] M R Trusheim E R Berndt and F L Douglas ldquoStratifiedmedicine strategic and economic implications of combiningdrugs and clinical biomarkersrdquoNature Reviews Drug Discoveryvol 6 no 4 pp 287ndash293 2007

4 BioMed Research International

(a) (b) (c)

(d) (e) (f)

(g) (h)

Figure 2 Example of an image of a CRC tissue section analyzed by EPSTRA (a) K8 staining (TxRed) overlaid with DAPI (blue) includingCy2 (green) and Cy5 (white) background fluorescence (b) epithelial mask (yellow) and stromal mask (green) as found by EPSTRA (c) theadjacent mucosa is depicted in green contours (d) tumor center in red (e) invasive front 1 in yellow and invasive front 2 in pink (f)ndash(h)show the corresponding epithelial mask of (c)ndash(e) views Bubbles show the region of interest (ROI) identification number and its area Barindicates 200120583m

(b) Some cell clusters containing very few cells (they mayrepresent tumor buds)

(c) Holes in the K8 grayscale channel due to lack of K8antigen inside the nuclear compartment

(d) Some nonspecific areas appearing more intenselystained than the positive K8 (erythrocytes auto-fluorescence lipofuscin folded tissue nonuniformtissue reactivity etc)

(e) Multiple intensity-classes of pixels each displayingmultiple peaks in the histogram lumen S weak Eintense E folded S folded E and erythrocytes

(f) Erythrocytes being intense in both TxRed and GFPchannels

(g) E areas which can be very close to each other Inthe center of the tumor the S may only form a verynarrow band between crypts (as thin as 5 pixels wide)

(h) Big data (up to 100000 images per slide each imagewith 14 mega-pixels)

The characteristics and novelty of the proposed method aregiven by the way of integrating the particularity of thesebiological signals (the four input channels and the intensities

BioMed Research International 5

IDAPI

ITexa

IFITC

ICy5

IeDAPI

IeTexa

IeFITC

IeCy5

VCtissue

VCepi

L mask

Tmask

Smask

Emask

AGSm(VCtissue1199831min1199831max)

AGSm(VCepi1199832

min1199832max) cap Tmask

Figure 3 Processing data flow

distribution for E S and lumen) into widely accepted fastimage preprocessing and segmentation methods

26 Region-Based Segmentation Approach Since our goalwas to find the boundaries for the areas of interest we usedregion-based segmentation based on Otsursquos thresholdingmethod [19] Let one consider that 119879ref = thrOtsu(119868) is thethreshold computed with this method on the entire inputimage 119868 Segmenting the entire image with the same globalthreshold (119879ref ) could cause the loss of some epithelial orstroma areas due to intrastaining variability Therefore weused an adaptive threshold applied on image tiles (119879

119894) as

provided by the TissueFAXS file formatA disadvantage of Otsu thresholding is that it separates

the image in two pixel classes even though the image mayhave only one class of pixels (eg stroma could be detectedwithin a tile only with lumen) In addition it may fail whenprovided with images with a multitude of the classes presentin the images (eg lumen weak stroma intense stroma andepithelium)Thus we implemented a restriction such that thethreshold is accepted only if it is within a predefined intervalthrOtsu(119879119894) isin [Tref minus 120572Tref + 120572] where 120572 is a constantWe chose the reference threshold as the one given by Otsuthresholding applied on the entire virtual slide

If the obtained threshold is outside the accepted restrictedinterval the process is repeated by considering only the pix-els within the predefined interval [TminTmax] empiricallyestablished after analyzing multiple image sets All the pixelswith the intensity smaller than Tmin and all the pixels withthe intensity higher thanTmax are eliminated Applying thistransformation before the Otsu thresholding guaranties theproper segmentation of the two classes of interest We namedthis approach ldquoAdaptive Guided Segmentationrdquo

AGS (119868TminTmax) (1)

Since the acquired images present artifacts a simple yetefficient enhancement technique was used after the back-ground is removed by subtracting an appropriate predefinedconstant (119888) the intense areas are further emphasized byraising values of the pixels to a certain power (119901 gt 1)empirically determined for each channel

119868119890 = (119868 minus 119888)119901 (2)

None of the required regions (lumen stroma and epithelialregion) can be accurately extracted from a single chan-nel due to aforementioned intensity variability Therefore

a combination of these channels was used to create morerobust ldquovirtual channelsrdquo (VC) Thus a mix of enhancedDAPI TxRed FITC and Cy5 was made for tissue detection(VCtissue) and one for epithelial detection (VCepi)

VCtissue = 119868119890DAPI + 119868119890GFP + 119868119890TxRed + 119868119890Cy5

VCepi = 119868119890TxRed minus 119868119890GFP minus 119868119890Cy5(3)

The algorithmworkflow needed two branches one for lumen(Lmask) versus tissue (Tmask) masks identification and one forstroma (Smask) versus epithelial (Emask) identification withinthe tissue mask (Tmask) The tissue-lumen detection phaseextracts the tissue and lumenmasks respectively by applyingAGS on VCtissue

LmaskTmask = AGS (VCtissueT1

minT1

max) (4)

Subsequently the epithelial and stromamasks were extractedusing AGS on the VCepi only within tissue mask Tmask

SmaskEmask = AGS (VCepithT2

minT2

max)⋂Tmask (5)

AGS generates irregular contours typical to binarizationmethods Therefore we smoothed the contours by applyinga series of closing and opening morphological operators [20]further referred to as AGSm Since larger smoothing filtersmay generate a better smoothing at the cost of losing detailswe preferred using filters not larger than 5 pixels

The final data flow of the implemented segmentationtechnique is presented in Figure 3

Due to the huge amount of data which needed to beprocessed the proposed algorithmsegmentation techniquewas implemented using a parallel approach making use ofthe independent computing units (cores) available on currentprocessors On a 16-core computer an improvement of 14-fold in the analysis speed was noticed compared to sequentialsingle-core approach reaching an average time of 3min persection

27 Measurements and Statistical Processing Masks of E andS areas were saved as images corresponding to original imagetiles StrataQuest 50 was used to create analysis projectsfrom the original images as well as from the newly createdmask files With the module ldquoTotal Area Measurementrdquo thecalculation was done for the areas of two masks for everysingle marked group of regions normal mucosa mucosa

6 BioMed Research International

(a) (b)

(c) (d)

Figure 4 120573-cateninK8 double staining in a colorectal cancer tissue section (a) Merged image 120573-catenin positive cells are visible in red andK8 positive cells are shown in green Nuclei are shown in light blue (all false colors) The same image is shown with individual fluorescencechannels in grey scale (b) K8 (c) 120573-catenin and (d) nuclear staining with DAPI Arrows with solid lines indicate cells with a translocation(cytoplasmic or nuclear) of 120573-catenin Arrows with dashed lines indicate small groups of cells still positive for K8 but no more expressing120573-catenin

adjacent mucosa tumor center invasive front 1 and invasivefront 2

Batch export functionality extracted the final measure-ments in a single Excel file containing measurements of areaof E and S as well as their ratio for each group of region oneach slide Finally the remaining statistics (averages standarddeviations) were calculated in Excel

3 Results

31 K8 as aMarker for Epithelial Cells in CRC Todifferentiatebetween E and S area in CRC sections we used the epithelialcell marker K8 [21] To verify that the expression of K8 isappropriate to label epithelial cells at various cancer stagesin all patient groups we performed on an additional labelingwith 120573-catenin on 5 test slides which were checked visuallyIn normal epithelial cells 120573-catenin is located on the cellularmembrane at the lateral sides of the cells [22] During cancerprogression 120573-catenin accumulates in the cytoplasm andhence enters the nucleus (see examples in Figures 4(a) and4(c))

However the expression of 120573-catenin is lost in someepithelial cells within the tumor (Figures 4(a) and 4(c)arrows with dashed lines) but a strong expression of K8 ismaintained in all epithelial cells within all areas Thus weconclude that K8 can be considered a reliable marker fornormal as well as for cancerous epithelial cells in the colon

32 Region Definition and Assessment of EPSTRA MaskResults StrataQuest can overlay any combination of chan-nels and masks in various colors and transparencies Figure 2shows examples of such overlays in theway theywere used fordefining and categorizing ROIs EPSTRA mask and contourresults could be also overlaid for visual assessment

33 Comparison of EPSTRA versus Results from HumanExperts and Other Automated Methods Evaluation of themethod was done on a test dataset consisting of 24 FOVsfrom all defined compartments and patient groups Whenassessing the capability of algorithms the ground truthis considered to be results of human experts [23] Tocompensate for interobserver variability we considered theopinion of three different human experts By including themseparately in the performance assessment we can comparethe results from each individual against the performance ofthe proposed EPSTRAmethodThe ground truth (GrTr) wascreated by averaging the manual mark-up (average GrTr)which assigned pixels from the entire test dataset to threemasks lumen S and E After eliminating the lumen maskpixels classified by the method and by the averaged GrTras Ewere considered as ldquotrue positivesrdquo (TP) Also we calculatedldquofalse positiverdquo (FP = pixels marked as A by the algorithm butnot by theGrTr) ldquofalse negativerdquo (FN=pixels notmarked as Eby the algorithmbutmarked by theGrTr) and ldquotrue negativerdquo(TN = pixels not marked as E by neither the algorithm

BioMed Research International 7

Table 2 Specificity precision recall and F1 score from three humanexperts (hExp 1ndash3) alternative automated methods (GT GT + MPLT LT + MP and MH) and EPSTRA

Methodexpert Specificity Precision Recall F1 scorehExp1 097 099 098 098hExp2 098 099 097 098hExp3 098 099 098 099GT 097 099 034 051GT + MP 098 099 030 047LT 093 098 075 085LT + MP 096 099 074 085MH 096 099 080 089EPSTRA 097 099 097 098GT global thresholding using cumulative histogram thresholdingMPmor-phologic processing LT local thresholding MH morphological hierarchy

nor by the GrTr) Precision recallsensitivity specificity andbalanced F1 score [24] were calculated as follows

Specificity = TN(TP + FP)

Precision = TP(TP + FP)

Recall = TP(TP + FN)

F1 = 2 lowast Precision lowast Recall(Precision + Recall)

(6)

As alternative automated workflows we implemented Otsuglobal thresholding (GT) and local thresholding variant byprocessing each image tile independently (LT) Both thesetwo binarization methods were enhanced with additionalmorphological postprocessing (MP) to eliminate the holessituated in the place of the nuclei (methods named GT+ MP and LT + MP resp) We also evaluated the newermethod shown in [25] using morphological hierarchy (MH)although it was designed for and tested on healthy tissue onlyConsidering the results of enumerated alternative methodswe calculated the specificity precision recall and F1 scoreData is summarized in Table 2

34 Evaluation of the ES Ratio in the Defined CompartmentsData on the epithelialstromal ratio generated by EPSTRA issummarized in Figure 5

In healthy colon sections the mean ratio of 115 fornormal mucosa (NM) indicates that the amount of E roughlyequals the amount of S This ratio is similar in the normalappearing mucosa (Mu) distant from tumor area in CRCsections (094 for patients with and 106 for patients withoutliver metastasis) In the mucosa adjacent (AM) to the tumorthe mean ratio is slightly higher (133 for patients with and148 for patients without liver metastasis) An even higheramount of E area was found within the tumor center with anES of 167 for patients with and 201 for patients without livermetastasis At the invasive front 1 (IF1) the ES in patients

minusLM+LM

NM Mu AM TC IF1 IF2000

050

100

150

200

250

300

350

Epith

elia

lstro

mal

ratio

NM Mu AM TC IF1 IF2

115106 148 201 169 012

094 133 167 115 012+LM

minusLM

Figure 5 Epithelialstromal ratio in different tissue compartments(normal mucosa (NM) in healthy patients normal appearingmucosa (Mu) distant from tumor adjacent mucosa (AM) tumorcenter (TC) invasive front 1 (IF1) and invasive front 2 (IF2)) incolorectal cancer patients with and without liver metastasis Valuesare given as means plusmn standard deviation

without liver metastasis was 169 and it was reduced in thatwith livermetastasis (115) As expected at the invasive front 2(IF2) where the tissue is disrupted and only single tumor cellsor small clamps of cells are spread in the stroma the ES ratiois very low (mean ratio of 012 for patients with and withoutliver metastasis resp)

4 Discussion

We established an automatic microscopic imaging method tomeasure the E and S area in paraffin-embedded section fromCRC patients The novel algorithms for the calculation of theES ratio offer significant improvements of the histologicalevaluation of immunofluorescent stained tumor sectionsIn comparison with the HampE staining identification ofepithelial cells is increased due to specific labelling of keratin8 Virtual microscopy overlaying techniques bring additionaladvantages when defining regions of interest based on bothmorphology and intensity of the stained markers

Image processing techniques are used mainly becausethey allow large scale statistical evaluation in addition to clas-sical eye screening As noticed in Table 2 the new EPSTRAmethods achieved better results than existing automaticapproaches almost reaching the performance of humanexperts The ldquoGTrdquo method succeeds in extracting morerelevant threshold values but fails in adapting to existinglocal variations of the sample (ie tumor center shows loweroverall expression levels) On the other side ldquoLTrdquo does notfind proper values in areaswithout lumen or stroma EPSTRAuses information from GT but then continues to adapt tolocal variations succeeding in proper segmentation despiteintrinsic expression variability In addition the fact that itcould find both thin S areas inside tumor centers and smallE structures (which may represent tumor buds) embedded inthe S made it superior when compared to all other evaluated

8 BioMed Research International

automated methods In addition it performed at least asgood as the worse human expert delineating contours in adigital slide with a digital pen at original resolution Sincescoring (quick visual qualitative assessment) is typically fasterand more error-prone than manual drawing of contours ourautomated method would perform better than such visualscoring procedures Morphological postprocessing typicallyhelps to remove very small holes or structures (up to 5pixels) However using MP for holesstructures bigger than5ndash10 pixels dramatically decreases contour accuracy and therecognition rate of thin S andor isolated E cells Theseproblems are also specific to MH [25] and will also appearin other approaches that are based on image downscalingHowever EPSTRA successfully avoided these shortcomingsby working with images at original resolution (pixel size =0323 120583m)

Using EPSTRA on the entire image dataset we foundan increase in E in the central part of the tumor as it wasexpected from the uncontrolled growth of E cells At thetumor margins (IF1) the ES ratio in CRC patients withliver metastasis is 32 lower than in patients without livermetastasis These findings suggest that the lower amount ofE at the tumor margin is related to metastasis and could becaused by epithelial-mesenchymal transition (EMT) of tumorcells A similar reduction of ES ratio was noticed in TCas well Thus patients with the reduced ES ratio may havea shorter progression-free survival time as suggested fromprevious visual assessment studies in colon and breast cancer[12ndash14]

In tumor with increases in S areas E clusters seem tobe less compact The invasive margin shows a rupture ofthe E structure and even budding events Single E cells orsmall E cell clusters are floating in the S This histomor-phological structure [26] at the tumor margin is alreadyproposed as a parameter to be included in predicting CRCprognosis [27] Since budding is difficult to assess visuallythe newly designed EPSTRA method proves to be ade-quate for quantification of this surrogate marker Althoughthe method requires more extensive immunostaining andsophisticated imaging than traditional visual histopathologyit offers important benefits It could provide more addi-tional information than simply counting the tumor budsFurthermore EPSTRA could allowmultiple biomarkers to bequantified selectively over specified cell types regardless oftheir abundance

5 Conclusion

In this study we developed the EPSTRA method for auto-mated compartment basedmorphometric analysis and deter-mination of the ES ratios in CRC sections as a means ofadvanced tissue cytometry We demonstrated that EPSTRAcan be used as a powerful tool for objective quantitativemeasurements of tumor structural development with a scal-able analysis time and reproducibility of data In the futureapplying this technique and elaborating more elaboratedscoring procedures together with additional IF markers mayextend diagnostic options leading to a fast and reliablecharacterization of tumor properties in individual patients

Thereby it may help enabling a better personalized treatmentand individual follow-up in the patientrsquos evaluation promot-ing an improvement of the efficacy of a certain therapeuticregimen [28]

Conflict of Interests

R Rogojanu is a PhD student at Medical University ViennaAustria and is working also for TissueGnostics GmbHVienna Austria as Head of Research amp Development

Authorsrsquo Contribution

Writing was coordinated by T Thalhammer and accom-plished by R Rogojanu C Smochina W Jager and GBises Section staining was optimized and performed by UThiem and G Bises Pathological inspection and clinical datawere provided by I Mesteri Characterization of samples andhighlighting of tumor areas were performed by G BisesAlgorithm development was done by R Rogojanu A Heindland A Seewald Validation of the results was done by AHeindl C Smochina and R Rogojanu Entire project wasdesigned and coordinated by I Ellinger T Thalhammer andG Bises

Acknowledgment

Financial support was given by the FFG Bridge-Project 818094

References

[1] C L Sawyers ldquoThe cancer biomarker problemrdquo Nature vol452 no 7187 pp 548ndash552 2008

[2] S Ogino P Lochhead A T Chan et al ldquoMolecular pathologicalepidemiology of epigenetics emerging integrative science toanalyze environment host and diseaserdquoModern Pathology vol26 no 4 pp 465ndash484 2013

[3] E J A Morris N J Maughan D Forman and P Quirke ldquoWhoto treat with adjuvant therapy in Dukes Bstage II colorectalcancer The need for high quality pathologyrdquo Gut vol 56 no10 pp 1419ndash1425 2007

[4] J R Jass ldquoMolecular heterogeneity of colorectal cancer impli-cations for cancer controlrdquo Surgical Oncology vol 16 supple-ment 1 pp S7ndashS9 2007

[5] I Zlobec and A Lugli ldquoPrognostic and predictive factors incolorectal cancerrdquo Postgraduate Medical Journal vol 84 no994 pp 403ndash411 2008

[6] M Allen and J Louise Jones ldquoJekyll and Hyde the role ofthe microenvironment on the progression of cancerrdquo Journal ofPathology vol 223 no 2 pp 162ndash176 2011

[7] E Karamitopoulou I Zlobec I Panayiotides et al ldquoSystematicanalysis of proteins from different signaling pathways in thetumor center and the invasive front of colorectal cancerrdquoHuman Pathology vol 42 no 12 pp 1888ndash1896 2011

[8] M Sund and R Kalluri ldquoTumor stroma derived biomarkers incancerrdquo Cancer and Metastasis Reviews vol 28 no 1-2 pp 177ndash183 2009

BioMed Research International 9

[9] M Scurr A Gallimore and A Godkin ldquoT cell subsets andcolorectal cancer discerning the good from the badrdquo CellularImmunology vol 279 no 1 pp 21ndash24 2012

[10] Q-W Zhang L Liu C-Y Gong et al ldquoPrognostic significanceof tumor-associated macrophages in solid tumor a meta-analysis of the literaturerdquo PLoS ONE vol 7 no 12 Article IDe50946 2012

[11] J Betge M J Pollheimer R A Lindtner et al ldquoIntramural andextramural vascular invasion in colorectal cancer prognosticsignificance and quality of pathology reportingrdquo Cancer vol118 no 3 pp 628ndash638 2012

[12] N P West M Dattani P McShane et al ldquoThe proportionof tumour cells is an independent predictor for survival incolorectal cancer patientsrdquo British Journal of Cancer vol 102no 10 pp 1519ndash1523 2010

[13] A Huijbers R A E M Tollenaar G W V Pelt et al ldquoTheproportion of tumor-stroma as a strong prognosticator for stageII and III colon cancer patients validation in the victor trialrdquoAnnals of Oncology vol 24 no 1 Article ID mds246 pp 179ndash185 2013

[14] E M de Kruijf J G H van Nes C J H van de Velde et alldquoTumor-stroma ratio in the primary tumor is a prognostic fac-tor in early breast cancer patients especially in triple-negativecarcinoma patientsrdquo Breast Cancer Research and Treatment vol125 no 3 pp 687ndash696 2011

[15] E F W Courrech Staal V T H B M Smit M-L F VanVelthuysen et al ldquoReproducibility and validation of tumourstroma ratio scoring on oesophageal adenocarcinoma biopsiesrdquoEuropean Journal of Cancer vol 47 no 3 pp 375ndash382 2011

[16] C A Rubio ldquoCell proliferation at the leading invasive frontof colonic carcinomas Preliminary observationsrdquo AnticancerResearch vol 26 no 3 pp 2275ndash2278 2006

[17] A Jung M Schrauder U Oswald et al ldquoThe invasion frontof human colorectal adenocarcinomas shows co-localization ofnuclear 120573-catenin cyclin D1 and p16INK4A and is a region oflow proliferationrdquo The American Journal of Pathology vol 159no 5 pp 1613ndash1617 2001

[18] M Ponz de Leon L Roncucci P Di Donato et al ldquoPatternof epithelial cell proliferation in colorectal mucosa of normalsubjects and of patients with adenomatous polyps or cancer ofthe large bowelrdquo Cancer Research vol 48 no 14 pp 4121ndash41261988

[19] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[20] R C Gonzales and R E Woods Digital Image ProcessingPrentice-Hall 3rd edition 2007

[21] V Karantza ldquoKeratins in health and cancer more than mereepithelial cell markersrdquo Oncogene vol 30 no 2 pp 127ndash1382011

[22] T Brabletz F Hlubek S Spaderna et al ldquoInvasion and metas-tasis in colorectal cancer epithelial-mesenchymal transitionmesenchymal-epithelial transition stemcells and beta-cateninrdquoCells Tissues Organs vol 179 no 1-2 pp 56ndash65 2005

[23] S B Wilson and R Emerson ldquoSpike detection a review andcomparison of algorithmsrdquo Clinical Neurophysiology vol 113no 12 pp 1873ndash1881 2002

[24] C J van Rijsbergen Information Retrieval Butterworths Lon-don UK 2nd edition 1979

[25] C Smochina V Manta and W Kropatsch ldquoCrypts detectionin microscopic images using hierarchical structuresrdquo PatternRecognition Letters vol 34 no 8 pp 934ndash941 2013

[26] J P Sleeman G Christofori R Fodde et al ldquoConcepts ofmetastasis in flux the stromal progression modelrdquo Seminars inCancer Biology vol 22 no 3 pp 174ndash186 2012

[27] A Lugli E Karamitopoulou and I Zlobec ldquoTumour buddinga promising parameter in colorectal cancerrdquo British Journal ofCancer vol 106 no 11 pp 1713ndash1717 2012

[28] M R Trusheim E R Berndt and F L Douglas ldquoStratifiedmedicine strategic and economic implications of combiningdrugs and clinical biomarkersrdquoNature Reviews Drug Discoveryvol 6 no 4 pp 287ndash293 2007

BioMed Research International 5

IDAPI

ITexa

IFITC

ICy5

IeDAPI

IeTexa

IeFITC

IeCy5

VCtissue

VCepi

L mask

Tmask

Smask

Emask

AGSm(VCtissue1199831min1199831max)

AGSm(VCepi1199832

min1199832max) cap Tmask

Figure 3 Processing data flow

distribution for E S and lumen) into widely accepted fastimage preprocessing and segmentation methods

26 Region-Based Segmentation Approach Since our goalwas to find the boundaries for the areas of interest we usedregion-based segmentation based on Otsursquos thresholdingmethod [19] Let one consider that 119879ref = thrOtsu(119868) is thethreshold computed with this method on the entire inputimage 119868 Segmenting the entire image with the same globalthreshold (119879ref ) could cause the loss of some epithelial orstroma areas due to intrastaining variability Therefore weused an adaptive threshold applied on image tiles (119879

119894) as

provided by the TissueFAXS file formatA disadvantage of Otsu thresholding is that it separates

the image in two pixel classes even though the image mayhave only one class of pixels (eg stroma could be detectedwithin a tile only with lumen) In addition it may fail whenprovided with images with a multitude of the classes presentin the images (eg lumen weak stroma intense stroma andepithelium)Thus we implemented a restriction such that thethreshold is accepted only if it is within a predefined intervalthrOtsu(119879119894) isin [Tref minus 120572Tref + 120572] where 120572 is a constantWe chose the reference threshold as the one given by Otsuthresholding applied on the entire virtual slide

If the obtained threshold is outside the accepted restrictedinterval the process is repeated by considering only the pix-els within the predefined interval [TminTmax] empiricallyestablished after analyzing multiple image sets All the pixelswith the intensity smaller than Tmin and all the pixels withthe intensity higher thanTmax are eliminated Applying thistransformation before the Otsu thresholding guaranties theproper segmentation of the two classes of interest We namedthis approach ldquoAdaptive Guided Segmentationrdquo

AGS (119868TminTmax) (1)

Since the acquired images present artifacts a simple yetefficient enhancement technique was used after the back-ground is removed by subtracting an appropriate predefinedconstant (119888) the intense areas are further emphasized byraising values of the pixels to a certain power (119901 gt 1)empirically determined for each channel

119868119890 = (119868 minus 119888)119901 (2)

None of the required regions (lumen stroma and epithelialregion) can be accurately extracted from a single chan-nel due to aforementioned intensity variability Therefore

a combination of these channels was used to create morerobust ldquovirtual channelsrdquo (VC) Thus a mix of enhancedDAPI TxRed FITC and Cy5 was made for tissue detection(VCtissue) and one for epithelial detection (VCepi)

VCtissue = 119868119890DAPI + 119868119890GFP + 119868119890TxRed + 119868119890Cy5

VCepi = 119868119890TxRed minus 119868119890GFP minus 119868119890Cy5(3)

The algorithmworkflow needed two branches one for lumen(Lmask) versus tissue (Tmask) masks identification and one forstroma (Smask) versus epithelial (Emask) identification withinthe tissue mask (Tmask) The tissue-lumen detection phaseextracts the tissue and lumenmasks respectively by applyingAGS on VCtissue

LmaskTmask = AGS (VCtissueT1

minT1

max) (4)

Subsequently the epithelial and stromamasks were extractedusing AGS on the VCepi only within tissue mask Tmask

SmaskEmask = AGS (VCepithT2

minT2

max)⋂Tmask (5)

AGS generates irregular contours typical to binarizationmethods Therefore we smoothed the contours by applyinga series of closing and opening morphological operators [20]further referred to as AGSm Since larger smoothing filtersmay generate a better smoothing at the cost of losing detailswe preferred using filters not larger than 5 pixels

The final data flow of the implemented segmentationtechnique is presented in Figure 3

Due to the huge amount of data which needed to beprocessed the proposed algorithmsegmentation techniquewas implemented using a parallel approach making use ofthe independent computing units (cores) available on currentprocessors On a 16-core computer an improvement of 14-fold in the analysis speed was noticed compared to sequentialsingle-core approach reaching an average time of 3min persection

27 Measurements and Statistical Processing Masks of E andS areas were saved as images corresponding to original imagetiles StrataQuest 50 was used to create analysis projectsfrom the original images as well as from the newly createdmask files With the module ldquoTotal Area Measurementrdquo thecalculation was done for the areas of two masks for everysingle marked group of regions normal mucosa mucosa

6 BioMed Research International

(a) (b)

(c) (d)

Figure 4 120573-cateninK8 double staining in a colorectal cancer tissue section (a) Merged image 120573-catenin positive cells are visible in red andK8 positive cells are shown in green Nuclei are shown in light blue (all false colors) The same image is shown with individual fluorescencechannels in grey scale (b) K8 (c) 120573-catenin and (d) nuclear staining with DAPI Arrows with solid lines indicate cells with a translocation(cytoplasmic or nuclear) of 120573-catenin Arrows with dashed lines indicate small groups of cells still positive for K8 but no more expressing120573-catenin

adjacent mucosa tumor center invasive front 1 and invasivefront 2

Batch export functionality extracted the final measure-ments in a single Excel file containing measurements of areaof E and S as well as their ratio for each group of region oneach slide Finally the remaining statistics (averages standarddeviations) were calculated in Excel

3 Results

31 K8 as aMarker for Epithelial Cells in CRC Todifferentiatebetween E and S area in CRC sections we used the epithelialcell marker K8 [21] To verify that the expression of K8 isappropriate to label epithelial cells at various cancer stagesin all patient groups we performed on an additional labelingwith 120573-catenin on 5 test slides which were checked visuallyIn normal epithelial cells 120573-catenin is located on the cellularmembrane at the lateral sides of the cells [22] During cancerprogression 120573-catenin accumulates in the cytoplasm andhence enters the nucleus (see examples in Figures 4(a) and4(c))

However the expression of 120573-catenin is lost in someepithelial cells within the tumor (Figures 4(a) and 4(c)arrows with dashed lines) but a strong expression of K8 ismaintained in all epithelial cells within all areas Thus weconclude that K8 can be considered a reliable marker fornormal as well as for cancerous epithelial cells in the colon

32 Region Definition and Assessment of EPSTRA MaskResults StrataQuest can overlay any combination of chan-nels and masks in various colors and transparencies Figure 2shows examples of such overlays in theway theywere used fordefining and categorizing ROIs EPSTRA mask and contourresults could be also overlaid for visual assessment

33 Comparison of EPSTRA versus Results from HumanExperts and Other Automated Methods Evaluation of themethod was done on a test dataset consisting of 24 FOVsfrom all defined compartments and patient groups Whenassessing the capability of algorithms the ground truthis considered to be results of human experts [23] Tocompensate for interobserver variability we considered theopinion of three different human experts By including themseparately in the performance assessment we can comparethe results from each individual against the performance ofthe proposed EPSTRAmethodThe ground truth (GrTr) wascreated by averaging the manual mark-up (average GrTr)which assigned pixels from the entire test dataset to threemasks lumen S and E After eliminating the lumen maskpixels classified by the method and by the averaged GrTras Ewere considered as ldquotrue positivesrdquo (TP) Also we calculatedldquofalse positiverdquo (FP = pixels marked as A by the algorithm butnot by theGrTr) ldquofalse negativerdquo (FN=pixels notmarked as Eby the algorithmbutmarked by theGrTr) and ldquotrue negativerdquo(TN = pixels not marked as E by neither the algorithm

BioMed Research International 7

Table 2 Specificity precision recall and F1 score from three humanexperts (hExp 1ndash3) alternative automated methods (GT GT + MPLT LT + MP and MH) and EPSTRA

Methodexpert Specificity Precision Recall F1 scorehExp1 097 099 098 098hExp2 098 099 097 098hExp3 098 099 098 099GT 097 099 034 051GT + MP 098 099 030 047LT 093 098 075 085LT + MP 096 099 074 085MH 096 099 080 089EPSTRA 097 099 097 098GT global thresholding using cumulative histogram thresholdingMPmor-phologic processing LT local thresholding MH morphological hierarchy

nor by the GrTr) Precision recallsensitivity specificity andbalanced F1 score [24] were calculated as follows

Specificity = TN(TP + FP)

Precision = TP(TP + FP)

Recall = TP(TP + FN)

F1 = 2 lowast Precision lowast Recall(Precision + Recall)

(6)

As alternative automated workflows we implemented Otsuglobal thresholding (GT) and local thresholding variant byprocessing each image tile independently (LT) Both thesetwo binarization methods were enhanced with additionalmorphological postprocessing (MP) to eliminate the holessituated in the place of the nuclei (methods named GT+ MP and LT + MP resp) We also evaluated the newermethod shown in [25] using morphological hierarchy (MH)although it was designed for and tested on healthy tissue onlyConsidering the results of enumerated alternative methodswe calculated the specificity precision recall and F1 scoreData is summarized in Table 2

34 Evaluation of the ES Ratio in the Defined CompartmentsData on the epithelialstromal ratio generated by EPSTRA issummarized in Figure 5

In healthy colon sections the mean ratio of 115 fornormal mucosa (NM) indicates that the amount of E roughlyequals the amount of S This ratio is similar in the normalappearing mucosa (Mu) distant from tumor area in CRCsections (094 for patients with and 106 for patients withoutliver metastasis) In the mucosa adjacent (AM) to the tumorthe mean ratio is slightly higher (133 for patients with and148 for patients without liver metastasis) An even higheramount of E area was found within the tumor center with anES of 167 for patients with and 201 for patients without livermetastasis At the invasive front 1 (IF1) the ES in patients

minusLM+LM

NM Mu AM TC IF1 IF2000

050

100

150

200

250

300

350

Epith

elia

lstro

mal

ratio

NM Mu AM TC IF1 IF2

115106 148 201 169 012

094 133 167 115 012+LM

minusLM

Figure 5 Epithelialstromal ratio in different tissue compartments(normal mucosa (NM) in healthy patients normal appearingmucosa (Mu) distant from tumor adjacent mucosa (AM) tumorcenter (TC) invasive front 1 (IF1) and invasive front 2 (IF2)) incolorectal cancer patients with and without liver metastasis Valuesare given as means plusmn standard deviation

without liver metastasis was 169 and it was reduced in thatwith livermetastasis (115) As expected at the invasive front 2(IF2) where the tissue is disrupted and only single tumor cellsor small clamps of cells are spread in the stroma the ES ratiois very low (mean ratio of 012 for patients with and withoutliver metastasis resp)

4 Discussion

We established an automatic microscopic imaging method tomeasure the E and S area in paraffin-embedded section fromCRC patients The novel algorithms for the calculation of theES ratio offer significant improvements of the histologicalevaluation of immunofluorescent stained tumor sectionsIn comparison with the HampE staining identification ofepithelial cells is increased due to specific labelling of keratin8 Virtual microscopy overlaying techniques bring additionaladvantages when defining regions of interest based on bothmorphology and intensity of the stained markers

Image processing techniques are used mainly becausethey allow large scale statistical evaluation in addition to clas-sical eye screening As noticed in Table 2 the new EPSTRAmethods achieved better results than existing automaticapproaches almost reaching the performance of humanexperts The ldquoGTrdquo method succeeds in extracting morerelevant threshold values but fails in adapting to existinglocal variations of the sample (ie tumor center shows loweroverall expression levels) On the other side ldquoLTrdquo does notfind proper values in areaswithout lumen or stroma EPSTRAuses information from GT but then continues to adapt tolocal variations succeeding in proper segmentation despiteintrinsic expression variability In addition the fact that itcould find both thin S areas inside tumor centers and smallE structures (which may represent tumor buds) embedded inthe S made it superior when compared to all other evaluated

8 BioMed Research International

automated methods In addition it performed at least asgood as the worse human expert delineating contours in adigital slide with a digital pen at original resolution Sincescoring (quick visual qualitative assessment) is typically fasterand more error-prone than manual drawing of contours ourautomated method would perform better than such visualscoring procedures Morphological postprocessing typicallyhelps to remove very small holes or structures (up to 5pixels) However using MP for holesstructures bigger than5ndash10 pixels dramatically decreases contour accuracy and therecognition rate of thin S andor isolated E cells Theseproblems are also specific to MH [25] and will also appearin other approaches that are based on image downscalingHowever EPSTRA successfully avoided these shortcomingsby working with images at original resolution (pixel size =0323 120583m)

Using EPSTRA on the entire image dataset we foundan increase in E in the central part of the tumor as it wasexpected from the uncontrolled growth of E cells At thetumor margins (IF1) the ES ratio in CRC patients withliver metastasis is 32 lower than in patients without livermetastasis These findings suggest that the lower amount ofE at the tumor margin is related to metastasis and could becaused by epithelial-mesenchymal transition (EMT) of tumorcells A similar reduction of ES ratio was noticed in TCas well Thus patients with the reduced ES ratio may havea shorter progression-free survival time as suggested fromprevious visual assessment studies in colon and breast cancer[12ndash14]

In tumor with increases in S areas E clusters seem tobe less compact The invasive margin shows a rupture ofthe E structure and even budding events Single E cells orsmall E cell clusters are floating in the S This histomor-phological structure [26] at the tumor margin is alreadyproposed as a parameter to be included in predicting CRCprognosis [27] Since budding is difficult to assess visuallythe newly designed EPSTRA method proves to be ade-quate for quantification of this surrogate marker Althoughthe method requires more extensive immunostaining andsophisticated imaging than traditional visual histopathologyit offers important benefits It could provide more addi-tional information than simply counting the tumor budsFurthermore EPSTRA could allowmultiple biomarkers to bequantified selectively over specified cell types regardless oftheir abundance

5 Conclusion

In this study we developed the EPSTRA method for auto-mated compartment basedmorphometric analysis and deter-mination of the ES ratios in CRC sections as a means ofadvanced tissue cytometry We demonstrated that EPSTRAcan be used as a powerful tool for objective quantitativemeasurements of tumor structural development with a scal-able analysis time and reproducibility of data In the futureapplying this technique and elaborating more elaboratedscoring procedures together with additional IF markers mayextend diagnostic options leading to a fast and reliablecharacterization of tumor properties in individual patients

Thereby it may help enabling a better personalized treatmentand individual follow-up in the patientrsquos evaluation promot-ing an improvement of the efficacy of a certain therapeuticregimen [28]

Conflict of Interests

R Rogojanu is a PhD student at Medical University ViennaAustria and is working also for TissueGnostics GmbHVienna Austria as Head of Research amp Development

Authorsrsquo Contribution

Writing was coordinated by T Thalhammer and accom-plished by R Rogojanu C Smochina W Jager and GBises Section staining was optimized and performed by UThiem and G Bises Pathological inspection and clinical datawere provided by I Mesteri Characterization of samples andhighlighting of tumor areas were performed by G BisesAlgorithm development was done by R Rogojanu A Heindland A Seewald Validation of the results was done by AHeindl C Smochina and R Rogojanu Entire project wasdesigned and coordinated by I Ellinger T Thalhammer andG Bises

Acknowledgment

Financial support was given by the FFG Bridge-Project 818094

References

[1] C L Sawyers ldquoThe cancer biomarker problemrdquo Nature vol452 no 7187 pp 548ndash552 2008

[2] S Ogino P Lochhead A T Chan et al ldquoMolecular pathologicalepidemiology of epigenetics emerging integrative science toanalyze environment host and diseaserdquoModern Pathology vol26 no 4 pp 465ndash484 2013

[3] E J A Morris N J Maughan D Forman and P Quirke ldquoWhoto treat with adjuvant therapy in Dukes Bstage II colorectalcancer The need for high quality pathologyrdquo Gut vol 56 no10 pp 1419ndash1425 2007

[4] J R Jass ldquoMolecular heterogeneity of colorectal cancer impli-cations for cancer controlrdquo Surgical Oncology vol 16 supple-ment 1 pp S7ndashS9 2007

[5] I Zlobec and A Lugli ldquoPrognostic and predictive factors incolorectal cancerrdquo Postgraduate Medical Journal vol 84 no994 pp 403ndash411 2008

[6] M Allen and J Louise Jones ldquoJekyll and Hyde the role ofthe microenvironment on the progression of cancerrdquo Journal ofPathology vol 223 no 2 pp 162ndash176 2011

[7] E Karamitopoulou I Zlobec I Panayiotides et al ldquoSystematicanalysis of proteins from different signaling pathways in thetumor center and the invasive front of colorectal cancerrdquoHuman Pathology vol 42 no 12 pp 1888ndash1896 2011

[8] M Sund and R Kalluri ldquoTumor stroma derived biomarkers incancerrdquo Cancer and Metastasis Reviews vol 28 no 1-2 pp 177ndash183 2009

BioMed Research International 9

[9] M Scurr A Gallimore and A Godkin ldquoT cell subsets andcolorectal cancer discerning the good from the badrdquo CellularImmunology vol 279 no 1 pp 21ndash24 2012

[10] Q-W Zhang L Liu C-Y Gong et al ldquoPrognostic significanceof tumor-associated macrophages in solid tumor a meta-analysis of the literaturerdquo PLoS ONE vol 7 no 12 Article IDe50946 2012

[11] J Betge M J Pollheimer R A Lindtner et al ldquoIntramural andextramural vascular invasion in colorectal cancer prognosticsignificance and quality of pathology reportingrdquo Cancer vol118 no 3 pp 628ndash638 2012

[12] N P West M Dattani P McShane et al ldquoThe proportionof tumour cells is an independent predictor for survival incolorectal cancer patientsrdquo British Journal of Cancer vol 102no 10 pp 1519ndash1523 2010

[13] A Huijbers R A E M Tollenaar G W V Pelt et al ldquoTheproportion of tumor-stroma as a strong prognosticator for stageII and III colon cancer patients validation in the victor trialrdquoAnnals of Oncology vol 24 no 1 Article ID mds246 pp 179ndash185 2013

[14] E M de Kruijf J G H van Nes C J H van de Velde et alldquoTumor-stroma ratio in the primary tumor is a prognostic fac-tor in early breast cancer patients especially in triple-negativecarcinoma patientsrdquo Breast Cancer Research and Treatment vol125 no 3 pp 687ndash696 2011

[15] E F W Courrech Staal V T H B M Smit M-L F VanVelthuysen et al ldquoReproducibility and validation of tumourstroma ratio scoring on oesophageal adenocarcinoma biopsiesrdquoEuropean Journal of Cancer vol 47 no 3 pp 375ndash382 2011

[16] C A Rubio ldquoCell proliferation at the leading invasive frontof colonic carcinomas Preliminary observationsrdquo AnticancerResearch vol 26 no 3 pp 2275ndash2278 2006

[17] A Jung M Schrauder U Oswald et al ldquoThe invasion frontof human colorectal adenocarcinomas shows co-localization ofnuclear 120573-catenin cyclin D1 and p16INK4A and is a region oflow proliferationrdquo The American Journal of Pathology vol 159no 5 pp 1613ndash1617 2001

[18] M Ponz de Leon L Roncucci P Di Donato et al ldquoPatternof epithelial cell proliferation in colorectal mucosa of normalsubjects and of patients with adenomatous polyps or cancer ofthe large bowelrdquo Cancer Research vol 48 no 14 pp 4121ndash41261988

[19] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[20] R C Gonzales and R E Woods Digital Image ProcessingPrentice-Hall 3rd edition 2007

[21] V Karantza ldquoKeratins in health and cancer more than mereepithelial cell markersrdquo Oncogene vol 30 no 2 pp 127ndash1382011

[22] T Brabletz F Hlubek S Spaderna et al ldquoInvasion and metas-tasis in colorectal cancer epithelial-mesenchymal transitionmesenchymal-epithelial transition stemcells and beta-cateninrdquoCells Tissues Organs vol 179 no 1-2 pp 56ndash65 2005

[23] S B Wilson and R Emerson ldquoSpike detection a review andcomparison of algorithmsrdquo Clinical Neurophysiology vol 113no 12 pp 1873ndash1881 2002

[24] C J van Rijsbergen Information Retrieval Butterworths Lon-don UK 2nd edition 1979

[25] C Smochina V Manta and W Kropatsch ldquoCrypts detectionin microscopic images using hierarchical structuresrdquo PatternRecognition Letters vol 34 no 8 pp 934ndash941 2013

[26] J P Sleeman G Christofori R Fodde et al ldquoConcepts ofmetastasis in flux the stromal progression modelrdquo Seminars inCancer Biology vol 22 no 3 pp 174ndash186 2012

[27] A Lugli E Karamitopoulou and I Zlobec ldquoTumour buddinga promising parameter in colorectal cancerrdquo British Journal ofCancer vol 106 no 11 pp 1713ndash1717 2012

[28] M R Trusheim E R Berndt and F L Douglas ldquoStratifiedmedicine strategic and economic implications of combiningdrugs and clinical biomarkersrdquoNature Reviews Drug Discoveryvol 6 no 4 pp 287ndash293 2007

6 BioMed Research International

(a) (b)

(c) (d)

Figure 4 120573-cateninK8 double staining in a colorectal cancer tissue section (a) Merged image 120573-catenin positive cells are visible in red andK8 positive cells are shown in green Nuclei are shown in light blue (all false colors) The same image is shown with individual fluorescencechannels in grey scale (b) K8 (c) 120573-catenin and (d) nuclear staining with DAPI Arrows with solid lines indicate cells with a translocation(cytoplasmic or nuclear) of 120573-catenin Arrows with dashed lines indicate small groups of cells still positive for K8 but no more expressing120573-catenin

adjacent mucosa tumor center invasive front 1 and invasivefront 2

Batch export functionality extracted the final measure-ments in a single Excel file containing measurements of areaof E and S as well as their ratio for each group of region oneach slide Finally the remaining statistics (averages standarddeviations) were calculated in Excel

3 Results

31 K8 as aMarker for Epithelial Cells in CRC Todifferentiatebetween E and S area in CRC sections we used the epithelialcell marker K8 [21] To verify that the expression of K8 isappropriate to label epithelial cells at various cancer stagesin all patient groups we performed on an additional labelingwith 120573-catenin on 5 test slides which were checked visuallyIn normal epithelial cells 120573-catenin is located on the cellularmembrane at the lateral sides of the cells [22] During cancerprogression 120573-catenin accumulates in the cytoplasm andhence enters the nucleus (see examples in Figures 4(a) and4(c))

However the expression of 120573-catenin is lost in someepithelial cells within the tumor (Figures 4(a) and 4(c)arrows with dashed lines) but a strong expression of K8 ismaintained in all epithelial cells within all areas Thus weconclude that K8 can be considered a reliable marker fornormal as well as for cancerous epithelial cells in the colon

32 Region Definition and Assessment of EPSTRA MaskResults StrataQuest can overlay any combination of chan-nels and masks in various colors and transparencies Figure 2shows examples of such overlays in theway theywere used fordefining and categorizing ROIs EPSTRA mask and contourresults could be also overlaid for visual assessment

33 Comparison of EPSTRA versus Results from HumanExperts and Other Automated Methods Evaluation of themethod was done on a test dataset consisting of 24 FOVsfrom all defined compartments and patient groups Whenassessing the capability of algorithms the ground truthis considered to be results of human experts [23] Tocompensate for interobserver variability we considered theopinion of three different human experts By including themseparately in the performance assessment we can comparethe results from each individual against the performance ofthe proposed EPSTRAmethodThe ground truth (GrTr) wascreated by averaging the manual mark-up (average GrTr)which assigned pixels from the entire test dataset to threemasks lumen S and E After eliminating the lumen maskpixels classified by the method and by the averaged GrTras Ewere considered as ldquotrue positivesrdquo (TP) Also we calculatedldquofalse positiverdquo (FP = pixels marked as A by the algorithm butnot by theGrTr) ldquofalse negativerdquo (FN=pixels notmarked as Eby the algorithmbutmarked by theGrTr) and ldquotrue negativerdquo(TN = pixels not marked as E by neither the algorithm

BioMed Research International 7

Table 2 Specificity precision recall and F1 score from three humanexperts (hExp 1ndash3) alternative automated methods (GT GT + MPLT LT + MP and MH) and EPSTRA

Methodexpert Specificity Precision Recall F1 scorehExp1 097 099 098 098hExp2 098 099 097 098hExp3 098 099 098 099GT 097 099 034 051GT + MP 098 099 030 047LT 093 098 075 085LT + MP 096 099 074 085MH 096 099 080 089EPSTRA 097 099 097 098GT global thresholding using cumulative histogram thresholdingMPmor-phologic processing LT local thresholding MH morphological hierarchy

nor by the GrTr) Precision recallsensitivity specificity andbalanced F1 score [24] were calculated as follows

Specificity = TN(TP + FP)

Precision = TP(TP + FP)

Recall = TP(TP + FN)

F1 = 2 lowast Precision lowast Recall(Precision + Recall)

(6)

As alternative automated workflows we implemented Otsuglobal thresholding (GT) and local thresholding variant byprocessing each image tile independently (LT) Both thesetwo binarization methods were enhanced with additionalmorphological postprocessing (MP) to eliminate the holessituated in the place of the nuclei (methods named GT+ MP and LT + MP resp) We also evaluated the newermethod shown in [25] using morphological hierarchy (MH)although it was designed for and tested on healthy tissue onlyConsidering the results of enumerated alternative methodswe calculated the specificity precision recall and F1 scoreData is summarized in Table 2

34 Evaluation of the ES Ratio in the Defined CompartmentsData on the epithelialstromal ratio generated by EPSTRA issummarized in Figure 5

In healthy colon sections the mean ratio of 115 fornormal mucosa (NM) indicates that the amount of E roughlyequals the amount of S This ratio is similar in the normalappearing mucosa (Mu) distant from tumor area in CRCsections (094 for patients with and 106 for patients withoutliver metastasis) In the mucosa adjacent (AM) to the tumorthe mean ratio is slightly higher (133 for patients with and148 for patients without liver metastasis) An even higheramount of E area was found within the tumor center with anES of 167 for patients with and 201 for patients without livermetastasis At the invasive front 1 (IF1) the ES in patients

minusLM+LM

NM Mu AM TC IF1 IF2000

050

100

150

200

250

300

350

Epith

elia

lstro

mal

ratio

NM Mu AM TC IF1 IF2

115106 148 201 169 012

094 133 167 115 012+LM

minusLM

Figure 5 Epithelialstromal ratio in different tissue compartments(normal mucosa (NM) in healthy patients normal appearingmucosa (Mu) distant from tumor adjacent mucosa (AM) tumorcenter (TC) invasive front 1 (IF1) and invasive front 2 (IF2)) incolorectal cancer patients with and without liver metastasis Valuesare given as means plusmn standard deviation

without liver metastasis was 169 and it was reduced in thatwith livermetastasis (115) As expected at the invasive front 2(IF2) where the tissue is disrupted and only single tumor cellsor small clamps of cells are spread in the stroma the ES ratiois very low (mean ratio of 012 for patients with and withoutliver metastasis resp)

4 Discussion

We established an automatic microscopic imaging method tomeasure the E and S area in paraffin-embedded section fromCRC patients The novel algorithms for the calculation of theES ratio offer significant improvements of the histologicalevaluation of immunofluorescent stained tumor sectionsIn comparison with the HampE staining identification ofepithelial cells is increased due to specific labelling of keratin8 Virtual microscopy overlaying techniques bring additionaladvantages when defining regions of interest based on bothmorphology and intensity of the stained markers

Image processing techniques are used mainly becausethey allow large scale statistical evaluation in addition to clas-sical eye screening As noticed in Table 2 the new EPSTRAmethods achieved better results than existing automaticapproaches almost reaching the performance of humanexperts The ldquoGTrdquo method succeeds in extracting morerelevant threshold values but fails in adapting to existinglocal variations of the sample (ie tumor center shows loweroverall expression levels) On the other side ldquoLTrdquo does notfind proper values in areaswithout lumen or stroma EPSTRAuses information from GT but then continues to adapt tolocal variations succeeding in proper segmentation despiteintrinsic expression variability In addition the fact that itcould find both thin S areas inside tumor centers and smallE structures (which may represent tumor buds) embedded inthe S made it superior when compared to all other evaluated

8 BioMed Research International

automated methods In addition it performed at least asgood as the worse human expert delineating contours in adigital slide with a digital pen at original resolution Sincescoring (quick visual qualitative assessment) is typically fasterand more error-prone than manual drawing of contours ourautomated method would perform better than such visualscoring procedures Morphological postprocessing typicallyhelps to remove very small holes or structures (up to 5pixels) However using MP for holesstructures bigger than5ndash10 pixels dramatically decreases contour accuracy and therecognition rate of thin S andor isolated E cells Theseproblems are also specific to MH [25] and will also appearin other approaches that are based on image downscalingHowever EPSTRA successfully avoided these shortcomingsby working with images at original resolution (pixel size =0323 120583m)

Using EPSTRA on the entire image dataset we foundan increase in E in the central part of the tumor as it wasexpected from the uncontrolled growth of E cells At thetumor margins (IF1) the ES ratio in CRC patients withliver metastasis is 32 lower than in patients without livermetastasis These findings suggest that the lower amount ofE at the tumor margin is related to metastasis and could becaused by epithelial-mesenchymal transition (EMT) of tumorcells A similar reduction of ES ratio was noticed in TCas well Thus patients with the reduced ES ratio may havea shorter progression-free survival time as suggested fromprevious visual assessment studies in colon and breast cancer[12ndash14]

In tumor with increases in S areas E clusters seem tobe less compact The invasive margin shows a rupture ofthe E structure and even budding events Single E cells orsmall E cell clusters are floating in the S This histomor-phological structure [26] at the tumor margin is alreadyproposed as a parameter to be included in predicting CRCprognosis [27] Since budding is difficult to assess visuallythe newly designed EPSTRA method proves to be ade-quate for quantification of this surrogate marker Althoughthe method requires more extensive immunostaining andsophisticated imaging than traditional visual histopathologyit offers important benefits It could provide more addi-tional information than simply counting the tumor budsFurthermore EPSTRA could allowmultiple biomarkers to bequantified selectively over specified cell types regardless oftheir abundance

5 Conclusion

In this study we developed the EPSTRA method for auto-mated compartment basedmorphometric analysis and deter-mination of the ES ratios in CRC sections as a means ofadvanced tissue cytometry We demonstrated that EPSTRAcan be used as a powerful tool for objective quantitativemeasurements of tumor structural development with a scal-able analysis time and reproducibility of data In the futureapplying this technique and elaborating more elaboratedscoring procedures together with additional IF markers mayextend diagnostic options leading to a fast and reliablecharacterization of tumor properties in individual patients

Thereby it may help enabling a better personalized treatmentand individual follow-up in the patientrsquos evaluation promot-ing an improvement of the efficacy of a certain therapeuticregimen [28]

Conflict of Interests

R Rogojanu is a PhD student at Medical University ViennaAustria and is working also for TissueGnostics GmbHVienna Austria as Head of Research amp Development

Authorsrsquo Contribution

Writing was coordinated by T Thalhammer and accom-plished by R Rogojanu C Smochina W Jager and GBises Section staining was optimized and performed by UThiem and G Bises Pathological inspection and clinical datawere provided by I Mesteri Characterization of samples andhighlighting of tumor areas were performed by G BisesAlgorithm development was done by R Rogojanu A Heindland A Seewald Validation of the results was done by AHeindl C Smochina and R Rogojanu Entire project wasdesigned and coordinated by I Ellinger T Thalhammer andG Bises

Acknowledgment

Financial support was given by the FFG Bridge-Project 818094

References

[1] C L Sawyers ldquoThe cancer biomarker problemrdquo Nature vol452 no 7187 pp 548ndash552 2008

[2] S Ogino P Lochhead A T Chan et al ldquoMolecular pathologicalepidemiology of epigenetics emerging integrative science toanalyze environment host and diseaserdquoModern Pathology vol26 no 4 pp 465ndash484 2013

[3] E J A Morris N J Maughan D Forman and P Quirke ldquoWhoto treat with adjuvant therapy in Dukes Bstage II colorectalcancer The need for high quality pathologyrdquo Gut vol 56 no10 pp 1419ndash1425 2007

[4] J R Jass ldquoMolecular heterogeneity of colorectal cancer impli-cations for cancer controlrdquo Surgical Oncology vol 16 supple-ment 1 pp S7ndashS9 2007

[5] I Zlobec and A Lugli ldquoPrognostic and predictive factors incolorectal cancerrdquo Postgraduate Medical Journal vol 84 no994 pp 403ndash411 2008

[6] M Allen and J Louise Jones ldquoJekyll and Hyde the role ofthe microenvironment on the progression of cancerrdquo Journal ofPathology vol 223 no 2 pp 162ndash176 2011

[7] E Karamitopoulou I Zlobec I Panayiotides et al ldquoSystematicanalysis of proteins from different signaling pathways in thetumor center and the invasive front of colorectal cancerrdquoHuman Pathology vol 42 no 12 pp 1888ndash1896 2011

[8] M Sund and R Kalluri ldquoTumor stroma derived biomarkers incancerrdquo Cancer and Metastasis Reviews vol 28 no 1-2 pp 177ndash183 2009

BioMed Research International 9

[9] M Scurr A Gallimore and A Godkin ldquoT cell subsets andcolorectal cancer discerning the good from the badrdquo CellularImmunology vol 279 no 1 pp 21ndash24 2012

[10] Q-W Zhang L Liu C-Y Gong et al ldquoPrognostic significanceof tumor-associated macrophages in solid tumor a meta-analysis of the literaturerdquo PLoS ONE vol 7 no 12 Article IDe50946 2012

[11] J Betge M J Pollheimer R A Lindtner et al ldquoIntramural andextramural vascular invasion in colorectal cancer prognosticsignificance and quality of pathology reportingrdquo Cancer vol118 no 3 pp 628ndash638 2012

[12] N P West M Dattani P McShane et al ldquoThe proportionof tumour cells is an independent predictor for survival incolorectal cancer patientsrdquo British Journal of Cancer vol 102no 10 pp 1519ndash1523 2010

[13] A Huijbers R A E M Tollenaar G W V Pelt et al ldquoTheproportion of tumor-stroma as a strong prognosticator for stageII and III colon cancer patients validation in the victor trialrdquoAnnals of Oncology vol 24 no 1 Article ID mds246 pp 179ndash185 2013

[14] E M de Kruijf J G H van Nes C J H van de Velde et alldquoTumor-stroma ratio in the primary tumor is a prognostic fac-tor in early breast cancer patients especially in triple-negativecarcinoma patientsrdquo Breast Cancer Research and Treatment vol125 no 3 pp 687ndash696 2011

[15] E F W Courrech Staal V T H B M Smit M-L F VanVelthuysen et al ldquoReproducibility and validation of tumourstroma ratio scoring on oesophageal adenocarcinoma biopsiesrdquoEuropean Journal of Cancer vol 47 no 3 pp 375ndash382 2011

[16] C A Rubio ldquoCell proliferation at the leading invasive frontof colonic carcinomas Preliminary observationsrdquo AnticancerResearch vol 26 no 3 pp 2275ndash2278 2006

[17] A Jung M Schrauder U Oswald et al ldquoThe invasion frontof human colorectal adenocarcinomas shows co-localization ofnuclear 120573-catenin cyclin D1 and p16INK4A and is a region oflow proliferationrdquo The American Journal of Pathology vol 159no 5 pp 1613ndash1617 2001

[18] M Ponz de Leon L Roncucci P Di Donato et al ldquoPatternof epithelial cell proliferation in colorectal mucosa of normalsubjects and of patients with adenomatous polyps or cancer ofthe large bowelrdquo Cancer Research vol 48 no 14 pp 4121ndash41261988

[19] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[20] R C Gonzales and R E Woods Digital Image ProcessingPrentice-Hall 3rd edition 2007

[21] V Karantza ldquoKeratins in health and cancer more than mereepithelial cell markersrdquo Oncogene vol 30 no 2 pp 127ndash1382011

[22] T Brabletz F Hlubek S Spaderna et al ldquoInvasion and metas-tasis in colorectal cancer epithelial-mesenchymal transitionmesenchymal-epithelial transition stemcells and beta-cateninrdquoCells Tissues Organs vol 179 no 1-2 pp 56ndash65 2005

[23] S B Wilson and R Emerson ldquoSpike detection a review andcomparison of algorithmsrdquo Clinical Neurophysiology vol 113no 12 pp 1873ndash1881 2002

[24] C J van Rijsbergen Information Retrieval Butterworths Lon-don UK 2nd edition 1979

[25] C Smochina V Manta and W Kropatsch ldquoCrypts detectionin microscopic images using hierarchical structuresrdquo PatternRecognition Letters vol 34 no 8 pp 934ndash941 2013

[26] J P Sleeman G Christofori R Fodde et al ldquoConcepts ofmetastasis in flux the stromal progression modelrdquo Seminars inCancer Biology vol 22 no 3 pp 174ndash186 2012

[27] A Lugli E Karamitopoulou and I Zlobec ldquoTumour buddinga promising parameter in colorectal cancerrdquo British Journal ofCancer vol 106 no 11 pp 1713ndash1717 2012

[28] M R Trusheim E R Berndt and F L Douglas ldquoStratifiedmedicine strategic and economic implications of combiningdrugs and clinical biomarkersrdquoNature Reviews Drug Discoveryvol 6 no 4 pp 287ndash293 2007

BioMed Research International 7

Table 2 Specificity precision recall and F1 score from three humanexperts (hExp 1ndash3) alternative automated methods (GT GT + MPLT LT + MP and MH) and EPSTRA

Methodexpert Specificity Precision Recall F1 scorehExp1 097 099 098 098hExp2 098 099 097 098hExp3 098 099 098 099GT 097 099 034 051GT + MP 098 099 030 047LT 093 098 075 085LT + MP 096 099 074 085MH 096 099 080 089EPSTRA 097 099 097 098GT global thresholding using cumulative histogram thresholdingMPmor-phologic processing LT local thresholding MH morphological hierarchy

nor by the GrTr) Precision recallsensitivity specificity andbalanced F1 score [24] were calculated as follows

Specificity = TN(TP + FP)

Precision = TP(TP + FP)

Recall = TP(TP + FN)

F1 = 2 lowast Precision lowast Recall(Precision + Recall)

(6)

As alternative automated workflows we implemented Otsuglobal thresholding (GT) and local thresholding variant byprocessing each image tile independently (LT) Both thesetwo binarization methods were enhanced with additionalmorphological postprocessing (MP) to eliminate the holessituated in the place of the nuclei (methods named GT+ MP and LT + MP resp) We also evaluated the newermethod shown in [25] using morphological hierarchy (MH)although it was designed for and tested on healthy tissue onlyConsidering the results of enumerated alternative methodswe calculated the specificity precision recall and F1 scoreData is summarized in Table 2

34 Evaluation of the ES Ratio in the Defined CompartmentsData on the epithelialstromal ratio generated by EPSTRA issummarized in Figure 5

In healthy colon sections the mean ratio of 115 fornormal mucosa (NM) indicates that the amount of E roughlyequals the amount of S This ratio is similar in the normalappearing mucosa (Mu) distant from tumor area in CRCsections (094 for patients with and 106 for patients withoutliver metastasis) In the mucosa adjacent (AM) to the tumorthe mean ratio is slightly higher (133 for patients with and148 for patients without liver metastasis) An even higheramount of E area was found within the tumor center with anES of 167 for patients with and 201 for patients without livermetastasis At the invasive front 1 (IF1) the ES in patients

minusLM+LM

NM Mu AM TC IF1 IF2000

050

100

150

200

250

300

350

Epith

elia

lstro

mal

ratio

NM Mu AM TC IF1 IF2

115106 148 201 169 012

094 133 167 115 012+LM

minusLM

Figure 5 Epithelialstromal ratio in different tissue compartments(normal mucosa (NM) in healthy patients normal appearingmucosa (Mu) distant from tumor adjacent mucosa (AM) tumorcenter (TC) invasive front 1 (IF1) and invasive front 2 (IF2)) incolorectal cancer patients with and without liver metastasis Valuesare given as means plusmn standard deviation

without liver metastasis was 169 and it was reduced in thatwith livermetastasis (115) As expected at the invasive front 2(IF2) where the tissue is disrupted and only single tumor cellsor small clamps of cells are spread in the stroma the ES ratiois very low (mean ratio of 012 for patients with and withoutliver metastasis resp)

4 Discussion

We established an automatic microscopic imaging method tomeasure the E and S area in paraffin-embedded section fromCRC patients The novel algorithms for the calculation of theES ratio offer significant improvements of the histologicalevaluation of immunofluorescent stained tumor sectionsIn comparison with the HampE staining identification ofepithelial cells is increased due to specific labelling of keratin8 Virtual microscopy overlaying techniques bring additionaladvantages when defining regions of interest based on bothmorphology and intensity of the stained markers

Image processing techniques are used mainly becausethey allow large scale statistical evaluation in addition to clas-sical eye screening As noticed in Table 2 the new EPSTRAmethods achieved better results than existing automaticapproaches almost reaching the performance of humanexperts The ldquoGTrdquo method succeeds in extracting morerelevant threshold values but fails in adapting to existinglocal variations of the sample (ie tumor center shows loweroverall expression levels) On the other side ldquoLTrdquo does notfind proper values in areaswithout lumen or stroma EPSTRAuses information from GT but then continues to adapt tolocal variations succeeding in proper segmentation despiteintrinsic expression variability In addition the fact that itcould find both thin S areas inside tumor centers and smallE structures (which may represent tumor buds) embedded inthe S made it superior when compared to all other evaluated

8 BioMed Research International

automated methods In addition it performed at least asgood as the worse human expert delineating contours in adigital slide with a digital pen at original resolution Sincescoring (quick visual qualitative assessment) is typically fasterand more error-prone than manual drawing of contours ourautomated method would perform better than such visualscoring procedures Morphological postprocessing typicallyhelps to remove very small holes or structures (up to 5pixels) However using MP for holesstructures bigger than5ndash10 pixels dramatically decreases contour accuracy and therecognition rate of thin S andor isolated E cells Theseproblems are also specific to MH [25] and will also appearin other approaches that are based on image downscalingHowever EPSTRA successfully avoided these shortcomingsby working with images at original resolution (pixel size =0323 120583m)

Using EPSTRA on the entire image dataset we foundan increase in E in the central part of the tumor as it wasexpected from the uncontrolled growth of E cells At thetumor margins (IF1) the ES ratio in CRC patients withliver metastasis is 32 lower than in patients without livermetastasis These findings suggest that the lower amount ofE at the tumor margin is related to metastasis and could becaused by epithelial-mesenchymal transition (EMT) of tumorcells A similar reduction of ES ratio was noticed in TCas well Thus patients with the reduced ES ratio may havea shorter progression-free survival time as suggested fromprevious visual assessment studies in colon and breast cancer[12ndash14]

In tumor with increases in S areas E clusters seem tobe less compact The invasive margin shows a rupture ofthe E structure and even budding events Single E cells orsmall E cell clusters are floating in the S This histomor-phological structure [26] at the tumor margin is alreadyproposed as a parameter to be included in predicting CRCprognosis [27] Since budding is difficult to assess visuallythe newly designed EPSTRA method proves to be ade-quate for quantification of this surrogate marker Althoughthe method requires more extensive immunostaining andsophisticated imaging than traditional visual histopathologyit offers important benefits It could provide more addi-tional information than simply counting the tumor budsFurthermore EPSTRA could allowmultiple biomarkers to bequantified selectively over specified cell types regardless oftheir abundance

5 Conclusion

In this study we developed the EPSTRA method for auto-mated compartment basedmorphometric analysis and deter-mination of the ES ratios in CRC sections as a means ofadvanced tissue cytometry We demonstrated that EPSTRAcan be used as a powerful tool for objective quantitativemeasurements of tumor structural development with a scal-able analysis time and reproducibility of data In the futureapplying this technique and elaborating more elaboratedscoring procedures together with additional IF markers mayextend diagnostic options leading to a fast and reliablecharacterization of tumor properties in individual patients

Thereby it may help enabling a better personalized treatmentand individual follow-up in the patientrsquos evaluation promot-ing an improvement of the efficacy of a certain therapeuticregimen [28]

Conflict of Interests

R Rogojanu is a PhD student at Medical University ViennaAustria and is working also for TissueGnostics GmbHVienna Austria as Head of Research amp Development

Authorsrsquo Contribution

Writing was coordinated by T Thalhammer and accom-plished by R Rogojanu C Smochina W Jager and GBises Section staining was optimized and performed by UThiem and G Bises Pathological inspection and clinical datawere provided by I Mesteri Characterization of samples andhighlighting of tumor areas were performed by G BisesAlgorithm development was done by R Rogojanu A Heindland A Seewald Validation of the results was done by AHeindl C Smochina and R Rogojanu Entire project wasdesigned and coordinated by I Ellinger T Thalhammer andG Bises

Acknowledgment

Financial support was given by the FFG Bridge-Project 818094

References

[1] C L Sawyers ldquoThe cancer biomarker problemrdquo Nature vol452 no 7187 pp 548ndash552 2008

[2] S Ogino P Lochhead A T Chan et al ldquoMolecular pathologicalepidemiology of epigenetics emerging integrative science toanalyze environment host and diseaserdquoModern Pathology vol26 no 4 pp 465ndash484 2013

[3] E J A Morris N J Maughan D Forman and P Quirke ldquoWhoto treat with adjuvant therapy in Dukes Bstage II colorectalcancer The need for high quality pathologyrdquo Gut vol 56 no10 pp 1419ndash1425 2007

[4] J R Jass ldquoMolecular heterogeneity of colorectal cancer impli-cations for cancer controlrdquo Surgical Oncology vol 16 supple-ment 1 pp S7ndashS9 2007

[5] I Zlobec and A Lugli ldquoPrognostic and predictive factors incolorectal cancerrdquo Postgraduate Medical Journal vol 84 no994 pp 403ndash411 2008

[6] M Allen and J Louise Jones ldquoJekyll and Hyde the role ofthe microenvironment on the progression of cancerrdquo Journal ofPathology vol 223 no 2 pp 162ndash176 2011

[7] E Karamitopoulou I Zlobec I Panayiotides et al ldquoSystematicanalysis of proteins from different signaling pathways in thetumor center and the invasive front of colorectal cancerrdquoHuman Pathology vol 42 no 12 pp 1888ndash1896 2011

[8] M Sund and R Kalluri ldquoTumor stroma derived biomarkers incancerrdquo Cancer and Metastasis Reviews vol 28 no 1-2 pp 177ndash183 2009

BioMed Research International 9

[9] M Scurr A Gallimore and A Godkin ldquoT cell subsets andcolorectal cancer discerning the good from the badrdquo CellularImmunology vol 279 no 1 pp 21ndash24 2012

[10] Q-W Zhang L Liu C-Y Gong et al ldquoPrognostic significanceof tumor-associated macrophages in solid tumor a meta-analysis of the literaturerdquo PLoS ONE vol 7 no 12 Article IDe50946 2012

[11] J Betge M J Pollheimer R A Lindtner et al ldquoIntramural andextramural vascular invasion in colorectal cancer prognosticsignificance and quality of pathology reportingrdquo Cancer vol118 no 3 pp 628ndash638 2012

[12] N P West M Dattani P McShane et al ldquoThe proportionof tumour cells is an independent predictor for survival incolorectal cancer patientsrdquo British Journal of Cancer vol 102no 10 pp 1519ndash1523 2010

[13] A Huijbers R A E M Tollenaar G W V Pelt et al ldquoTheproportion of tumor-stroma as a strong prognosticator for stageII and III colon cancer patients validation in the victor trialrdquoAnnals of Oncology vol 24 no 1 Article ID mds246 pp 179ndash185 2013

[14] E M de Kruijf J G H van Nes C J H van de Velde et alldquoTumor-stroma ratio in the primary tumor is a prognostic fac-tor in early breast cancer patients especially in triple-negativecarcinoma patientsrdquo Breast Cancer Research and Treatment vol125 no 3 pp 687ndash696 2011

[15] E F W Courrech Staal V T H B M Smit M-L F VanVelthuysen et al ldquoReproducibility and validation of tumourstroma ratio scoring on oesophageal adenocarcinoma biopsiesrdquoEuropean Journal of Cancer vol 47 no 3 pp 375ndash382 2011

[16] C A Rubio ldquoCell proliferation at the leading invasive frontof colonic carcinomas Preliminary observationsrdquo AnticancerResearch vol 26 no 3 pp 2275ndash2278 2006

[17] A Jung M Schrauder U Oswald et al ldquoThe invasion frontof human colorectal adenocarcinomas shows co-localization ofnuclear 120573-catenin cyclin D1 and p16INK4A and is a region oflow proliferationrdquo The American Journal of Pathology vol 159no 5 pp 1613ndash1617 2001

[18] M Ponz de Leon L Roncucci P Di Donato et al ldquoPatternof epithelial cell proliferation in colorectal mucosa of normalsubjects and of patients with adenomatous polyps or cancer ofthe large bowelrdquo Cancer Research vol 48 no 14 pp 4121ndash41261988

[19] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[20] R C Gonzales and R E Woods Digital Image ProcessingPrentice-Hall 3rd edition 2007

[21] V Karantza ldquoKeratins in health and cancer more than mereepithelial cell markersrdquo Oncogene vol 30 no 2 pp 127ndash1382011

[22] T Brabletz F Hlubek S Spaderna et al ldquoInvasion and metas-tasis in colorectal cancer epithelial-mesenchymal transitionmesenchymal-epithelial transition stemcells and beta-cateninrdquoCells Tissues Organs vol 179 no 1-2 pp 56ndash65 2005

[23] S B Wilson and R Emerson ldquoSpike detection a review andcomparison of algorithmsrdquo Clinical Neurophysiology vol 113no 12 pp 1873ndash1881 2002

[24] C J van Rijsbergen Information Retrieval Butterworths Lon-don UK 2nd edition 1979

[25] C Smochina V Manta and W Kropatsch ldquoCrypts detectionin microscopic images using hierarchical structuresrdquo PatternRecognition Letters vol 34 no 8 pp 934ndash941 2013

[26] J P Sleeman G Christofori R Fodde et al ldquoConcepts ofmetastasis in flux the stromal progression modelrdquo Seminars inCancer Biology vol 22 no 3 pp 174ndash186 2012

[27] A Lugli E Karamitopoulou and I Zlobec ldquoTumour buddinga promising parameter in colorectal cancerrdquo British Journal ofCancer vol 106 no 11 pp 1713ndash1717 2012

[28] M R Trusheim E R Berndt and F L Douglas ldquoStratifiedmedicine strategic and economic implications of combiningdrugs and clinical biomarkersrdquoNature Reviews Drug Discoveryvol 6 no 4 pp 287ndash293 2007

8 BioMed Research International

automated methods In addition it performed at least asgood as the worse human expert delineating contours in adigital slide with a digital pen at original resolution Sincescoring (quick visual qualitative assessment) is typically fasterand more error-prone than manual drawing of contours ourautomated method would perform better than such visualscoring procedures Morphological postprocessing typicallyhelps to remove very small holes or structures (up to 5pixels) However using MP for holesstructures bigger than5ndash10 pixels dramatically decreases contour accuracy and therecognition rate of thin S andor isolated E cells Theseproblems are also specific to MH [25] and will also appearin other approaches that are based on image downscalingHowever EPSTRA successfully avoided these shortcomingsby working with images at original resolution (pixel size =0323 120583m)

Using EPSTRA on the entire image dataset we foundan increase in E in the central part of the tumor as it wasexpected from the uncontrolled growth of E cells At thetumor margins (IF1) the ES ratio in CRC patients withliver metastasis is 32 lower than in patients without livermetastasis These findings suggest that the lower amount ofE at the tumor margin is related to metastasis and could becaused by epithelial-mesenchymal transition (EMT) of tumorcells A similar reduction of ES ratio was noticed in TCas well Thus patients with the reduced ES ratio may havea shorter progression-free survival time as suggested fromprevious visual assessment studies in colon and breast cancer[12ndash14]

In tumor with increases in S areas E clusters seem tobe less compact The invasive margin shows a rupture ofthe E structure and even budding events Single E cells orsmall E cell clusters are floating in the S This histomor-phological structure [26] at the tumor margin is alreadyproposed as a parameter to be included in predicting CRCprognosis [27] Since budding is difficult to assess visuallythe newly designed EPSTRA method proves to be ade-quate for quantification of this surrogate marker Althoughthe method requires more extensive immunostaining andsophisticated imaging than traditional visual histopathologyit offers important benefits It could provide more addi-tional information than simply counting the tumor budsFurthermore EPSTRA could allowmultiple biomarkers to bequantified selectively over specified cell types regardless oftheir abundance

5 Conclusion

In this study we developed the EPSTRA method for auto-mated compartment basedmorphometric analysis and deter-mination of the ES ratios in CRC sections as a means ofadvanced tissue cytometry We demonstrated that EPSTRAcan be used as a powerful tool for objective quantitativemeasurements of tumor structural development with a scal-able analysis time and reproducibility of data In the futureapplying this technique and elaborating more elaboratedscoring procedures together with additional IF markers mayextend diagnostic options leading to a fast and reliablecharacterization of tumor properties in individual patients

Thereby it may help enabling a better personalized treatmentand individual follow-up in the patientrsquos evaluation promot-ing an improvement of the efficacy of a certain therapeuticregimen [28]

Conflict of Interests

R Rogojanu is a PhD student at Medical University ViennaAustria and is working also for TissueGnostics GmbHVienna Austria as Head of Research amp Development

Authorsrsquo Contribution

Writing was coordinated by T Thalhammer and accom-plished by R Rogojanu C Smochina W Jager and GBises Section staining was optimized and performed by UThiem and G Bises Pathological inspection and clinical datawere provided by I Mesteri Characterization of samples andhighlighting of tumor areas were performed by G BisesAlgorithm development was done by R Rogojanu A Heindland A Seewald Validation of the results was done by AHeindl C Smochina and R Rogojanu Entire project wasdesigned and coordinated by I Ellinger T Thalhammer andG Bises

Acknowledgment

Financial support was given by the FFG Bridge-Project 818094

References

[1] C L Sawyers ldquoThe cancer biomarker problemrdquo Nature vol452 no 7187 pp 548ndash552 2008

[2] S Ogino P Lochhead A T Chan et al ldquoMolecular pathologicalepidemiology of epigenetics emerging integrative science toanalyze environment host and diseaserdquoModern Pathology vol26 no 4 pp 465ndash484 2013

[3] E J A Morris N J Maughan D Forman and P Quirke ldquoWhoto treat with adjuvant therapy in Dukes Bstage II colorectalcancer The need for high quality pathologyrdquo Gut vol 56 no10 pp 1419ndash1425 2007

[4] J R Jass ldquoMolecular heterogeneity of colorectal cancer impli-cations for cancer controlrdquo Surgical Oncology vol 16 supple-ment 1 pp S7ndashS9 2007

[5] I Zlobec and A Lugli ldquoPrognostic and predictive factors incolorectal cancerrdquo Postgraduate Medical Journal vol 84 no994 pp 403ndash411 2008

[6] M Allen and J Louise Jones ldquoJekyll and Hyde the role ofthe microenvironment on the progression of cancerrdquo Journal ofPathology vol 223 no 2 pp 162ndash176 2011

[7] E Karamitopoulou I Zlobec I Panayiotides et al ldquoSystematicanalysis of proteins from different signaling pathways in thetumor center and the invasive front of colorectal cancerrdquoHuman Pathology vol 42 no 12 pp 1888ndash1896 2011

[8] M Sund and R Kalluri ldquoTumor stroma derived biomarkers incancerrdquo Cancer and Metastasis Reviews vol 28 no 1-2 pp 177ndash183 2009

BioMed Research International 9

[9] M Scurr A Gallimore and A Godkin ldquoT cell subsets andcolorectal cancer discerning the good from the badrdquo CellularImmunology vol 279 no 1 pp 21ndash24 2012

[10] Q-W Zhang L Liu C-Y Gong et al ldquoPrognostic significanceof tumor-associated macrophages in solid tumor a meta-analysis of the literaturerdquo PLoS ONE vol 7 no 12 Article IDe50946 2012

[11] J Betge M J Pollheimer R A Lindtner et al ldquoIntramural andextramural vascular invasion in colorectal cancer prognosticsignificance and quality of pathology reportingrdquo Cancer vol118 no 3 pp 628ndash638 2012

[12] N P West M Dattani P McShane et al ldquoThe proportionof tumour cells is an independent predictor for survival incolorectal cancer patientsrdquo British Journal of Cancer vol 102no 10 pp 1519ndash1523 2010

[13] A Huijbers R A E M Tollenaar G W V Pelt et al ldquoTheproportion of tumor-stroma as a strong prognosticator for stageII and III colon cancer patients validation in the victor trialrdquoAnnals of Oncology vol 24 no 1 Article ID mds246 pp 179ndash185 2013

[14] E M de Kruijf J G H van Nes C J H van de Velde et alldquoTumor-stroma ratio in the primary tumor is a prognostic fac-tor in early breast cancer patients especially in triple-negativecarcinoma patientsrdquo Breast Cancer Research and Treatment vol125 no 3 pp 687ndash696 2011

[15] E F W Courrech Staal V T H B M Smit M-L F VanVelthuysen et al ldquoReproducibility and validation of tumourstroma ratio scoring on oesophageal adenocarcinoma biopsiesrdquoEuropean Journal of Cancer vol 47 no 3 pp 375ndash382 2011

[16] C A Rubio ldquoCell proliferation at the leading invasive frontof colonic carcinomas Preliminary observationsrdquo AnticancerResearch vol 26 no 3 pp 2275ndash2278 2006

[17] A Jung M Schrauder U Oswald et al ldquoThe invasion frontof human colorectal adenocarcinomas shows co-localization ofnuclear 120573-catenin cyclin D1 and p16INK4A and is a region oflow proliferationrdquo The American Journal of Pathology vol 159no 5 pp 1613ndash1617 2001

[18] M Ponz de Leon L Roncucci P Di Donato et al ldquoPatternof epithelial cell proliferation in colorectal mucosa of normalsubjects and of patients with adenomatous polyps or cancer ofthe large bowelrdquo Cancer Research vol 48 no 14 pp 4121ndash41261988

[19] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[20] R C Gonzales and R E Woods Digital Image ProcessingPrentice-Hall 3rd edition 2007

[21] V Karantza ldquoKeratins in health and cancer more than mereepithelial cell markersrdquo Oncogene vol 30 no 2 pp 127ndash1382011

[22] T Brabletz F Hlubek S Spaderna et al ldquoInvasion and metas-tasis in colorectal cancer epithelial-mesenchymal transitionmesenchymal-epithelial transition stemcells and beta-cateninrdquoCells Tissues Organs vol 179 no 1-2 pp 56ndash65 2005

[23] S B Wilson and R Emerson ldquoSpike detection a review andcomparison of algorithmsrdquo Clinical Neurophysiology vol 113no 12 pp 1873ndash1881 2002

[24] C J van Rijsbergen Information Retrieval Butterworths Lon-don UK 2nd edition 1979

[25] C Smochina V Manta and W Kropatsch ldquoCrypts detectionin microscopic images using hierarchical structuresrdquo PatternRecognition Letters vol 34 no 8 pp 934ndash941 2013

[26] J P Sleeman G Christofori R Fodde et al ldquoConcepts ofmetastasis in flux the stromal progression modelrdquo Seminars inCancer Biology vol 22 no 3 pp 174ndash186 2012

[27] A Lugli E Karamitopoulou and I Zlobec ldquoTumour buddinga promising parameter in colorectal cancerrdquo British Journal ofCancer vol 106 no 11 pp 1713ndash1717 2012

[28] M R Trusheim E R Berndt and F L Douglas ldquoStratifiedmedicine strategic and economic implications of combiningdrugs and clinical biomarkersrdquoNature Reviews Drug Discoveryvol 6 no 4 pp 287ndash293 2007

BioMed Research International 9

[9] M Scurr A Gallimore and A Godkin ldquoT cell subsets andcolorectal cancer discerning the good from the badrdquo CellularImmunology vol 279 no 1 pp 21ndash24 2012

[10] Q-W Zhang L Liu C-Y Gong et al ldquoPrognostic significanceof tumor-associated macrophages in solid tumor a meta-analysis of the literaturerdquo PLoS ONE vol 7 no 12 Article IDe50946 2012

[11] J Betge M J Pollheimer R A Lindtner et al ldquoIntramural andextramural vascular invasion in colorectal cancer prognosticsignificance and quality of pathology reportingrdquo Cancer vol118 no 3 pp 628ndash638 2012

[12] N P West M Dattani P McShane et al ldquoThe proportionof tumour cells is an independent predictor for survival incolorectal cancer patientsrdquo British Journal of Cancer vol 102no 10 pp 1519ndash1523 2010

[13] A Huijbers R A E M Tollenaar G W V Pelt et al ldquoTheproportion of tumor-stroma as a strong prognosticator for stageII and III colon cancer patients validation in the victor trialrdquoAnnals of Oncology vol 24 no 1 Article ID mds246 pp 179ndash185 2013

[14] E M de Kruijf J G H van Nes C J H van de Velde et alldquoTumor-stroma ratio in the primary tumor is a prognostic fac-tor in early breast cancer patients especially in triple-negativecarcinoma patientsrdquo Breast Cancer Research and Treatment vol125 no 3 pp 687ndash696 2011

[15] E F W Courrech Staal V T H B M Smit M-L F VanVelthuysen et al ldquoReproducibility and validation of tumourstroma ratio scoring on oesophageal adenocarcinoma biopsiesrdquoEuropean Journal of Cancer vol 47 no 3 pp 375ndash382 2011

[16] C A Rubio ldquoCell proliferation at the leading invasive frontof colonic carcinomas Preliminary observationsrdquo AnticancerResearch vol 26 no 3 pp 2275ndash2278 2006

[17] A Jung M Schrauder U Oswald et al ldquoThe invasion frontof human colorectal adenocarcinomas shows co-localization ofnuclear 120573-catenin cyclin D1 and p16INK4A and is a region oflow proliferationrdquo The American Journal of Pathology vol 159no 5 pp 1613ndash1617 2001

[18] M Ponz de Leon L Roncucci P Di Donato et al ldquoPatternof epithelial cell proliferation in colorectal mucosa of normalsubjects and of patients with adenomatous polyps or cancer ofthe large bowelrdquo Cancer Research vol 48 no 14 pp 4121ndash41261988

[19] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[20] R C Gonzales and R E Woods Digital Image ProcessingPrentice-Hall 3rd edition 2007

[21] V Karantza ldquoKeratins in health and cancer more than mereepithelial cell markersrdquo Oncogene vol 30 no 2 pp 127ndash1382011

[22] T Brabletz F Hlubek S Spaderna et al ldquoInvasion and metas-tasis in colorectal cancer epithelial-mesenchymal transitionmesenchymal-epithelial transition stemcells and beta-cateninrdquoCells Tissues Organs vol 179 no 1-2 pp 56ndash65 2005

[23] S B Wilson and R Emerson ldquoSpike detection a review andcomparison of algorithmsrdquo Clinical Neurophysiology vol 113no 12 pp 1873ndash1881 2002

[24] C J van Rijsbergen Information Retrieval Butterworths Lon-don UK 2nd edition 1979

[25] C Smochina V Manta and W Kropatsch ldquoCrypts detectionin microscopic images using hierarchical structuresrdquo PatternRecognition Letters vol 34 no 8 pp 934ndash941 2013

[26] J P Sleeman G Christofori R Fodde et al ldquoConcepts ofmetastasis in flux the stromal progression modelrdquo Seminars inCancer Biology vol 22 no 3 pp 174ndash186 2012

[27] A Lugli E Karamitopoulou and I Zlobec ldquoTumour buddinga promising parameter in colorectal cancerrdquo British Journal ofCancer vol 106 no 11 pp 1713ndash1717 2012

[28] M R Trusheim E R Berndt and F L Douglas ldquoStratifiedmedicine strategic and economic implications of combiningdrugs and clinical biomarkersrdquoNature Reviews Drug Discoveryvol 6 no 4 pp 287ndash293 2007