European Journal of Cancer (2014) 50, 544– 552

A v a i l a b l e a t w w w . s c i e nc e d i r e c t . c o m

ScienceDirect

jour na l homepage : www.e jcancer . com

Comparison of visual and automated assessmentof tumour inflammatory infiltrates in patientswith colorectal cancer

0959-8049/$ - see front matter � 2013 Elsevier Ltd. All rights reserved.

http://dx.doi.org/10.1016/j.ejca.2013.11.003

⇑ Corresponding author: Address: Academic Department of Surgery, School of Medicine, University of Glasgow, Glasgow Royal InfiGlasgow G31 2ER, UK. Tel.: +44 0141 211 5435; fax: +44 0141 552 3229.

E-mail addresses: graemeguthrie@doctors.org.uk, Graeme.Guthrie@glasgow.ac.uk (G.J.K. Guthrie).1 These authors contributed equally to this work.

R. Forrest a,1, G.J.K. Guthrie a,⇑,1, C. Orange b, P.G. Horgan a, D.C. McMillan a,C.S.D. Roxburgh a

a Academic Unit of Surgery, School of Medicine, University of Glasgow, Royal Infirmary, Glasgow G31 2ER, UKb University Department of Pathology, Southern General Hospital, Glasgow, UK

Available online 11 December 2013

KEYWORDS

Tumour inflammatorycell infiltrateAutomated assessmentColorectal cancer

Abstract Background: Cancer-associated inflammation is increasingly recognised to be animportant determinant of oncological outcome. In colorectal cancer, the presence of peri-tumoural inflammatory/lymphocytic infiltrates predicts improved survival. To date, theseinfiltrates, assessed visually on haematoxylin and eosin (H&E) stained sections, have failedto enter routine clinical practice, partly due to their subjective assessment and considerableinter-observer variation. The present study aims to develop an automated scoring methodto enable consistent and reproducible assessment of tumour inflammatory infiltrates in colo-rectal cancer.Methods: 154 colorectal cancer patients who underwent curative resection were included inthe study. The local inflammatory infiltrate was assessed using the method described byKlintrup–Makinen. H&E tumour sections were uploaded to an image analysis programme(Slidepath, Leica Biosystems). An image analysis algorithm was developed to count theinflammatory cells at the invasive margin. The manual and automated assessments of thetumour inflammatory infiltrates were then compared.Results: The automated inflammatory cell counts assessed using the freehand annotationmethod (p < 0.001) and the rectangular box method (p < 0.001) were significantly associatedwith both K–M score (p < 0.001) and K–M grade (p < 0.001). The inflammatory cell countswere divided using quartiles to group tumours with similar inflammatory cell densities. Therewas good agreement between the manual and automated scoring methods (intraclass correla-tion coefficient (ICC) = 0.82). Similar to the visual K–M scoring system, the automated K–Mclassification of the inflammatory cell counts, using quartiles, was significantly associated with

rmary,

R. Forrest et al. / European Journal of Cancer 50 (2014) 544–552 545

venous invasion (p < 0.05) and modified Glasgow Prognostic Score (mGPS) (p 6 0.05). Onunivariate survival analysis, both automated K–M category (p < 0.05) and automated K–Mgrade (p < 0.005) were associated with cancer-specific survival.Conclusion: The results of the present study demonstrate that automated assessmenteffectively recapitulates the clinical value of visual assessment of the local inflammatory cellinfiltrate at the invasive margin of colorectal tumours. In addition, it is possible to obtainan objective assessment of tumour inflammatory infiltrates using routinely stained H&Esections. An automated, computer-based scoring method is therefore a workable and cost-effective approach to clinical assessment of local immune cell infiltrates in colorectal cancer.

� 2013 Elsevier Ltd. All rights reserved.

1. Introduction

Colorectal cancer is the second most common causeof cancer death in both men and women in the UnitedKingdom (UK) with 16,000 deaths per year (Cancer-stats, UK, 2010). Despite improvements in treatment,outcomes remain poor with approximately half of thoseundergoing curative resection dying from the disease [1].

In recent years it has become increasingly clear thatcancer-associated inflammation, in the form of localand systemic inflammatory responses, is a key determi-nant of progression and survival in colorectal cancer.In particular, there is consistent evidence that the pres-ence of a high grade local inflammatory cell infiltrateboth within the tumour and in the immediate microenvi-ronment predicts survival independent of tumour stagein colorectal cancer [2–4]. Many studies have reportedthat increasing density of inflammatory cells in andaround the tumour is associated with improved outcomein patients with colorectal cancer and this is thought torepresent the host anti-tumour response [4]. Further,there is good evidence that the immune classificationof tumours has independent and superior prognosticvalue when compared to traditional staging methods [3].

Despite the strong evidence supporting the prognosticvalue of inflammatory cell infiltrates, and the existence ofwell-described methods for the semi-quantitative assess-ment of inflammatory cell infiltration [2,3], the extent ofthe local inflammatory cell infiltrate is not routinelyconsidered in clinical practice and conventional stagingsystems such as tumour-node-metastasis (TNM) stageremain the mainstay in clinical practice. Reasons for thisinclude the complexity and lack of reproducibility ofscoring the inflammatory cell infiltrate caused by differ-ences in immunohistochemical staining methods betweendifferent units, the different cell types present and impor-tantly, the subjectivity of assessing the tumour inflamma-tory cell infiltrate.

Therefore a reliable and accurate measure of thetumour inflammatory cell infiltrate may be useful inthe refinement of staging the host inflammatoryresponse in clinical pathological practice.

There is now image analysis software capable ofpoint-scoring cells in routinely processed haematoxylinand eosin (H&E) tumour sections. Recent studies have

reported that computer-aided analysis has significantadvantages over manual scoring methods including:objectivity, accuracy and reproducibility [5–8]. There-fore, this modality may offer a method of standardisingthe assessment of the local inflammatory cell infiltrate inpatients with colorectal cancer.

The aim of the present study was to compare visualand automated assessment of tumour inflammatory cellinfiltration in patients with colorectal cancer.

2. Patients and Methods

Patients with colorectal cancer who, on the basis ofpre-operative staging and laparotomy findings, wereconsidered to have undergone an elective, potentiallycurative resection of colorectal cancer between 1997and 2006 in a single surgical unit at Glasgow RoyalInfirmary were included in the study. Tumours werestaged using the conventional tumour-node-metastasis(TNM) staging system, 7th Edition, 2010 [9]. Patientswith conditions known to elicit an acute or chronic sys-temic inflammatory response were excluded. These werenamely (i) pre-operative chemoradiotherapy, (ii) clinicalevidence of active pre-operative infection, or (iii) chronicactive inflammatory diseases such as rheumatoid arthri-tis. The study was approved by the Research EthicsCommittee, Glasgow Royal Infirmary, Glasgow.

2.1. Visual assessment of tumour inflammatory cell

infiltration

Assessment of the tumour inflammatory cell infiltratewas performed on original haematoxylin and eosin-stained full-sections of the tumour, considered to be rep-resentative of the specimen. The local inflammatoryresponse was evaluated previously in this cohort (GJKGand CSDR) using the method described by Klintrupet al. [2].

Briefly, K–M criteria is a four-point scale, a score of 0indicates no increase in inflammatory cells at the inva-sive margin, a score of 1 indicates presence of a mild/patchy increase in inflammatory cell reaction at theinvasive margin but no destruction of invading cancercell islets, a score of 2 indicates observation of a band-like inflammatory reaction at the invasive margin and

546 R. Forrest et al. / European Journal of Cancer 50 (2014) 544–552

a score of 3 indicates observation of a florid inflamma-tory reaction with cup-like inflammatory infiltrate atthe invasive margin [2].

The manual scores were then dichotomised to ‘high’and ‘low’ grade inflammation in line with the previouslypublished literature [2].

2.2. Slide scanning and automated assessment

The routine haematoxylin and eosin-stained tumoursections used for the visual assessment were scannedusing a Hamamatsu NanoZoomer (Welwyn GardenCity, Hertfordshire, UK). Visualisation and image anal-ysis assessment were carried out using Slidepath DigitalImage Hub, version 4.0.1 (Slidepath, Leica Biosystems,Milton Keynes, UK). Visual assessment of inflamma-tory infiltrates at the deepest point of the invasive mar-gin was performed on a high definition monitor. Threeslides were annotated and scored for each specimenand a total of 160 tumour specimens were scored.

The ‘Measure stained cells algorithm’, Tissue ImageAnalysis, version 2.0 (Slidepath, Leica Biosystems,Milton Keynes, UK), was then used to assess thesections for immune cell infiltrates at the point felt torepresent the deepest point of tumour invasion. Thealgorithm quantifies nuclear staining to derive a numericscore for each selected sample area. The default algo-rithm preferences were modified to count only inflam-matory cells at the invasive margin and exclude othercell types including stromal fibroblasts and tumour cells.This distinction between different cell types was basedon different staining intensities, cell size and size ofnuclei. The optimised algorithm parameters are shownin Supplementary data, Table 1.

Visual validation of the algorithm was performed toensure that only inflammatory cells were counted andother cell types were excluded from the analysis. Theanalysis scale was performed at 20� magnification forall scoring.

In order for the software system to be guided to ana-lyse the inflammatory infiltrate at the invasive margin itwas necessary to annotate the H&E sections. Two differ-ent methods of annotation were used and their accuracyassessed.

These included:

1. Freehand annotations� Using the sealed freehand annotation tool pro-

vided by Slidepath software. The area to be ana-lysed was manually selected by drawing aroundthe inflammatory infiltrate at the invasive margin.All freehand annotations were drawn at 20�magnification.

� One freehand annotation was created for eachslide.

� The optimised algorithm was selected and thealgorithm customised to analyse only the anno-tated region at the invasive margin.

� Upon completion of automatic cell counting, man-ual export of the data in .csv format allowed forfurther analysis.

� Cell counts were expressed as total number ofpositive nuclei (Alm) and the number of positivestained cells/mm2. The total number of positivenuclei was obtained from the Slidepath data out-put and the number of cells/mm2 was calculatedby dividing the mean total number of cells by themean total tissue area analysed (mm2).

� Three slides were scored for each tumour specimenand the mean score of the slides was taken as thefinal score for that specimen.

2. Rectangular boxes� The area from the invasive margin to be analysed

was identified visually and the magnification wasthen set to 20� at the selected area.

� The ‘Tissue IA optimiser’ icon was selected and theoptimised ‘Measure stained cells algorithm’ waschosen to analyse the area of the invasive marginvisible on-screen. This generated rectangular boxesto indicate the position of the area analysed.Results were recorded and stored for each rectan-gular area.

� Cell counts were expressed as the total number ofpositive nuclei (Alm) and number of positivestained cells/mm2. The total number of positivenuclei was obtained from the Slidepath data out-put and the number of cells/mm2 was calculatedby dividing the mean total number of cells by themean total tissue area analysed (mm2).

� To ensure reproducibility of measurements, threerectangles were assessed along the invasive marginof each tumour using this method. The mean scoreof the three boxes was taken as the final score forthat slide.

� Three slides were scored for each tumour specimenand the mean score of the slides was taken as thefinal score for that specimen.

The tissue IA optimiser provides a means of analys-ing regions of interest in H&E slides in an objective,quantitative and reproducible manner through manipu-lation of programmable algorithms. This software toolhas been adopted in the present study to provide cellcounts to reflect the density of inflammatory cells atthe invasive margin. It is plausible that other softwareprogrammes providing cell quantification based on col-our separation per pixel level would have a similar capa-bility with minimal manipulation.

R. Forrest et al. / European Journal of Cancer 50 (2014) 544–552 547

2.3. Statistical analysis

Associations between visual K–M scores and auto-mated inflammatory cell counts were examined usingboxplots. An analysis of variance between the auto-mated cell counts associated with each K–M score wereperformed using the Kruskal–Wallis test. The inflamma-tory cell counts were categorised into groups using quar-tiles. All other variables were grouped according tostandard or previously published thresholds. Associa-tions between automated scoring and tumour variableswere examined using Chi-square tests for trend. Kap-lan–Meier survival curves with log-rank tests were usedto perform univariate survival analyses. A p-value < 0.05 was considered statistically significant. Sta-tistical analyses were performed using SPSS� version19.0 (IBM SPSS, Chicago, Illinois, United States ofAmerica (USA)).

3. Results

One hundred and thirty-four patients who underwentpotentially curative resection of colorectal cancer wereincluded in the study. Summary characteristics ofpatients included in the study are shown in Table 1.The majority of patients were 65 years or older (70%)with similar numbers of men (52%) and women (48%).The majority of resections were carried out electively(87%) and were for colon cancer (78%). Pathologicalreports classified the tumours as T-stage 1 (1%), 2(3%), 3 (72%), 4 (24%). On routine pathological analy-sis, the minority of patients had evidence of ‘high-risk’pathological features including poor tumour differentia-tion (10%), extramural vascular invasion (31%), perito-neal involvement (22%), margin involvement (6%) and

Table 1The clinicopathological characteristics of patients with primar

Clinicopathological characteristics

Age (<65/65–74/>75 years)

Sex (male/female)

Tumour characteristicsTumour site (colon/rectum)T-stage (T1/T2/T3/T4)N-stage (N0/N1/N2)Tumour-node-metastasis (TNM) stageDifferentiation (poor/moderate–well)Venous invasion (absent/present)Peritoneal involvement (absent/present)Margin involvement (absent/present)

Systemic responsesModified Glasgow prognostic score (0/1/2)Neutrophil–lymphocyte ratio (<5/>5)

Local inflammatory cell infiltrateVisual Klintrup–Makinen grade (0/1/2/3)Visual Klintrup–Makinen category (weak/strong)

therefore the majority were classed as ‘low’ PetersenIndex (88%). On visual assessment using Klintrup’s cri-teria, the majority of patients (69%) were given a scoreof 0 or 1 (low-grade inflammation) and the remainder(31%) were given a score of 2 or 3 (high-gradeinflammation).

The relationships between K–M scoring and theautomated inflammatory cell counts were examinedusing boxplots. The automated inflammatory cell countsassessed using the freehand annotation method were sig-nificantly associated with both K–M score (p < 0.001)and K–M grade (p < 0.001). The automated inflamma-tory cell counts assessed using the rectangular boxmethod were also significantly associated with both K–M score (p < 0.001) and K–M grade (p < 0.001). Themedian numbers of inflammatory cells associated witheach K–M score and each K–M grade are shown inTable 2, Figs. 1a and 1b.

An analysis of variance was performed to assess forsignificant differences in the median number of inflam-matory cells associated with each K–M score. The free-hand annotation method demonstrated significantdifferences in the number of inflammatory cells betweenall K–M scores, with the exception of scores 2 and 3.

The rectangular box method demonstrated significantdifferences in the number of inflammatory cells betweenall K–M scores, with the exception of the number ofcells per mm2 between scores 2 and 3 (Table 2). Whilstboth methods demonstrated significant variation acrossK–M categories, there was greater discrimination usingthe rectangular box method in comparison to the free-hand annotation method (Chi-square 86.2 versus 52.6).Therefore, the total number of inflammatory cells atthe invasive margin assessed using the rectangular boxesmethod was selected for further survival analysis.

y operable colorectal cancer. (n = 134).

Patients (n = 134)

40 (30%)/52 (39%)/42 (31%)

70 (52%)/64 (48%)

105 (78%)/29 (22%)1 (1%)/4 (3%)/97 (72%)/32 (24%)109 (81%)/16 (12%)/9 (7%)4 (3%)/105 (78%)/25 (19%)13 (10%)/121 (90%)93 (69%)/41 (31%)105 (78%)/29 (22%)126 (94%)/8 (6%)

80 (60%)/32 (24%)/22 (16%)80 (60%)/26 (19%)

37 (28%)/55 (41%)/26 (19%)/16 (12%)92 (69%)/42 (31%)

Table 2Median number of inflammatory cells associated with K–M scoring.

Totalinflammatorycells (median andrange)

Totalinflammatorycells (p-value)

Inflammatory cellsper mm2 (medianand range)

Inflammatorycells per mm2

(p-value)

Freehand annotation method

Manual K–M score0 670 (13–6524) 0.75 (0–3.65)1 1814 (17–7906) <0.001a 2.02 (0–4.75) <0.001e

2 3115 (1222–6806) <0.001b 2.86 (1.49–4.47) <0.001f

3 3107 (671–6483) 0.425c 3.06 (0.89–5.36) 0.690g

Manual K–M gradeWeak 1489 (13–7906) 1.64 (0.3–4.75)Strong 3107 (671–6806) <0.001d 2.9 (0.89–5.36) <0.001

Rectangular box method

Manual K–M score0 91 (0–583) 0.45 (0–3.38)1 346 (2–1043) <0.001a 1.81 (0.02–4.97) <0.001e

2 682 (369–1068) <0.001b 3.61 (1.46–5.6) <0.001f

3 852 (145–1227) <0.05c 3.95 (0.76–5.9) <0.347g

Manual K–M gradeWeak 273 (0–1043) 1.27 (0–4.97)Strong 742 (145–1227) <0.001d 3.7 (0.76–5.96) <0.001d

Freehand annotation method: Analysis of variance using Kruskal–Wallis test between:a Groups 0 and 1 (total inflammatory cells).b Groups 1 and 2 (total inflammatory cells).c Groups 2 and 3 (total inflammatory cells).d Analysis of variance between counts in weak and strong category (total inflammatory cells, and inflammatory cells per mm2).e Groups 0 and 1 (inflammatory cells per mm2).f Groups 1 and 2 (inflammatory cells per mm2).g Groups 2 and 3 (inflammatory cells per mm2).Rectangular box method: Analysis of variance using Kruskal–Wallis test between:a Groups 0 and 1 (total inflammatory cells).b Groups 1 and 2 (total inflammatory cells).c Groups 2 and 3 (total inflammatory cells).d Analysis of variance between counts in weak and strong category (- total inflammatory cells, and - inflammatory cells per mm2).e Groups 0 and 1 (inflammatory cells per mm2).f Groups 1 and 2 (inflammatory cells per mm2).g Groups 2 and 3 (inflammatory cells per mm2).

548 R. Forrest et al. / European Journal of Cancer 50 (2014) 544–552

The inflammatory cell counts were then divided intocategories to group tumours with similar inflammatorycell densities. Given that there were significant differ-ences in the number of inflammatory cells associatedwith each K–M score, using the rectangular boxesmethod, the automated cell counts were grouped intofour separate categories using quartiles. To test howstrongly the automated Klintrup quartiles were associ-ated with the visually scored K–M scores the intraclasscorrelation coefficient (ICC) was calculated and therewas good agreement with an ICC of 0.82.

Similar to the visual K–M scoring system, the auto-mated K–M classification of the inflammatory cellcounts, using quartiles, was significantly associated withvenous invasion (p < 0.05), and mGPS (p 6 0.05),Tables 3a and 3b. An assessment of the prognostic valueof the automated classification of inflammatory cellcounts was then performed. The median follow-up was

107 months (range 44–178). During this period, 43patients died from colorectal cancer and 38 patients diedfrom other causes.

The relationship between baseline clinicopathologicalcharacteristics and cancer-specific survival is shown inTable 4.

On univariate analysis, age (p < 0.05), T-stage(p < 0.005), N-stage (p < 0.005), TNM stage (p < 0.005),venous invasion (p < 0.05), peritoneal involvement(p < 0.05), margin involvement (p < 0.005), mGPS(p < 0.005) and visual K–M category (p < 0.005) andvisual K–M grade (p < 0.005) were significantly associ-ated with cancer-specific survival. Further, on univariatesurvival analysis, both automated K–M category(p < 0.05) and automated K–M grade (p < 0.005) wereassociated with cancer-specific survival.

Kaplan–Meier survival analysis using the manuallyacquired scores demonstrated that patients whose

Fig. 1a. Box-plot demonstrating the comparison between automatedcounting of cells per mm2 (freehand annotation) and visual assessmentof tumour inflammatory infiltrates using the Klintrup–Makinenmethod.

Fig. 1b. Box-plot demonstrating the comparison between automatedcounting of cells per mm2 (rectangular box annotation) and visualassessment of tumour inflammatory infiltrates using the Klintrup–Makinen method.

R. Forrest et al. / European Journal of Cancer 50 (2014) 544–552 549

tumours had high-grade inflammatory/immune cell infil-trate had improved cancer-specific survival comparedwith those with low-grade inflammatory cell infiltrate(Figs. 2a and 2b) as previously reported by Klintrupand colleagues. Similarly, on Kaplan–Meier survivalanalysis, both automated Klintrup category (p < 0.05)and automated Klintrup grade (p < 0.05) demonstratedthat those patients with high inflammatory cell countshad improved cancer-specific survival when comparedwith those patients with lower inflammatory cell counts(Figs. 3a and 3b).

4. Discussion

The results of the present study demonstrate thatusing, commercially available image analysis software,an automated objective assessment of the peritumouralinflammatory cell infiltrate at the invasive margin wasobtained and that such an assessment had prognosticvalue in patients with colorectal cancer.

In order for assessment of the local inflammatoryresponse to be adopted in routine clinical pathologicalpractice, the method used must be practical, reproduc-ible and accurate on routinely stained H&E tumoursections. In the present study, whilst both automatedmethods, rectangular boxes and freehand annotations,demonstrated significant variation across K–Mcategories, there was greater discrimination using therectangular boxes method. The rectangular box methodallowed more precise and selective sampling of inflam-matory infiltrates along the tumour invasive margin.The freehand method required sampling from a larger,continuous area of the invasive margin, therefore thesesamples were likely to include more background ‘noise’making it difficult to accurately point score the inflam-matory cells. In addition, the freehand annotations werenot of a standardised shape or size.

In the present small study, the rectangular box auto-mated scoring method had a tendency to over-score thelocal inflammatory cell infiltrate compared to the visualK–M scoring system (4/47 graded as K–M high gradewere scored by automated method as ‘low inflammatorycell count’ and 15/93 graded as K–M low grade werescored by the automated method as ‘high inflammatorycell count’). These discrepancies are likely to be relatedto the algorithm as selection of valid algorithm parame-ters is of utmost importance. The algorithm used in thepresent study distinguished inflammatory cells fromother cell types found at the invasive margin, includingstromal fibroblasts and tumour cells.

This distinction was based on different stainingintensities, cell size and size of nuclei, amongst otherparameters and the algorithm was optimised to workadequately across all tumour specimens.

In the present study, the good agreement observedbetween the visual K–M scoring method and the auto-mated inflammatory cell counts support the accuracyof the image analysis scoring method for the assessmentof the local inflammatory response. In addition, andproviding further validation of accuracy, the automatedinflammatory cell counts were associated with a numberof pathological characteristics including T-stage, TNMstage, venous invasion, peritoneal and margin involve-ment. A key advantage of the automated method is thatit provides reliable numeric quantification of the inflam-matory cell density and thus provides an objective mea-sure that may also provide an improved degree ofaccuracy and sensitivity over the more subjective visual

Table 3aRelationships between automated Klintrup–Makinen (aKM) score using rectangular box method and clinicopathological variables.

Quartiles

0 (n = 30) 1 (n = 32) 2 (n = 37) 3 (n = 35) p-Value

Patient factorsAge (<65/65–74/P75 years) 6/12/12 10/15/7 9/14/14 15/11/9 0.158Sex (male/female) 19/11 12/20 19/18 20/15 0.972

Tumour characteristicsTumour site (colon/rectum) 27/3 23/9 28/9 27/8 0.312T-stage (T1/T2/T3/T4) 0/1/19/10 0/0/25/7 0/1/30/6 1/2/23/9 0.172N-stage (N0/N1/N2) 26/1/3 24/6/2 31/5/1 28/4/3 0.949Tumour-node-metastasis (TNM) stage 1/25/4 0/24/8 1/30/6 2/26/7 0.990Differentiation (poor/moderate–well) 5/25 2/30 2/35 4/31 0.515Venous invasion (absent/present) 15/15 24/8 27/10 27/8 0.033Peritoneal involvement (absent/present) 22/8 25/7 31/6 27/8 0.613Margin involvement (absent/present) 27/3 29/3 36/1 34/1 0.130

Systemic responsesModified Glasgow prognostic score (0/1/2) 11/10/9 24/5/3 23/7/7 22/10/3 0.05Neutrophil–lymphocyte ratio (<5:1/>5:1) 20/8 19/6 21/6 20/6 0.617

Table 3bRelationships between automated Klintrup–Makinen (aKM) category using rectangular box method and clinicopathological variables.

Automated Klintrup–Makinen score (aKM)

Weak Strong p-Value0 (n = 62) 1 (n = 72)

Patient factorsAge (<65/65–74/P75 years) 16/27/19 24/25/23 0.647Sex (male/female) 31/31 39/33 0.631

Tumour characteristicsTumour site (colon/rectum) 50/12 55/17 0.552T-stage (T1/T2/T3/T4) 0/1/44/17 1/3/53/15 0.180N-stage (N0/N1/N2) 50/7/5 59/9/4 0.700Tumour-node-metastasis (TNM) stage 1/49/12 3/56/13 0.613Differentiation (poor/moderate–well) 7/55 6/66 0.566Venous invasion (absent/present) 39/23 54/18 0.131Peritoneal involvement (absent/present) 47/15 58/14 0.507Margin involvement (absent/present) 56/6 70/2 0.094

Systemic responsesModified Glasgow prognostic score (0/1/2) 35/15/12 46/17/10 0.382Neutrophil–lymphocyte ratio (<5:1/>5:1) 39/14 41/12 0.653

550 R. Forrest et al. / European Journal of Cancer 50 (2014) 544–552

method. Indeed, several authors have proposed thatautomated assessment of inflammatory infiltrates offersadvantages over visual assessment including, consis-tency, exactness, objectivity and time-efficiency [7,8,10].

Previous studies examining the role of computer-aided assessment of tumour inflammatory infiltrateshave focused on the use automated cell counting of imu-nohistochemically stained tumour sections [7,8,10,11].However, very few immunohistochemical stains are usedin routine clinical practice, therefore the use of suchstains to aid automated cell counting adds a layer ofcomplexity to automated assessment that may precludeits use in routine clinical pathological practice in mostinstitutions.

While the H&E protocol used in our institution hasbeen consistent during the study period, we recognisethat other centres may have slightly different staining

protocols that may result in different staining intensitiesthat may affect the reproducibility of this method if thesame algorithm was to be applied to other cohorts.

While we acknowledge this limitation, we proposethat minor adjustment of the image analysis algorithmmay be able to allow for this. An automated scoringmethod will not replace the need for a human observerto determine that the images analysed are representativeof the lesion, do not have significant background stain-ing, and ensure areas of necrosis are avoided but willprovide a degree of consistency in observation not pos-sible with purely visual assessment. In addition, thecomputer-based scoring method used in this study takesadvantage of existing software available and is not a‘custom’ system, and is therefore easy to use and rela-tively inexpensive and could easily be acquired bypathology departments for routine use.

Table 4Baseline clinicopathological characteristics and relationships with cancer-specific survival (n = 134).

Cancer-specific survival

Univariate analysis

HR 95% CI p-Value

Patient factorsAge (<65/65–74/>75) 1.58 (1.01–2.47) 0.046Sex (male/female) 0.87 (0.44–1.71) 0.688

Tumour factorsTumour site (colon/rectum) 1.57 (0.75–3.29) 0.228T-stage (T1/T2/T3/T4) 2.55 (1.32–4.94) 0.005N-stage (N0/N1/N2) 2.04 (1.31–3.16) 0.002Tumour-node-metastasis (TNM) stage (I/II/III) 3.15 (1.59–6.25) 0.001Differentiation (poor/moderate–well) 1.45 (0.51–4.12) 0.489Venous invasion (absent/present) 2.30 (1.16–4.56) 0.017Peritoneal involvement (absent/present) 2.12 (1.03–4.38) 0.042Margin involvement (absent/present) 4.16 (1.60–10.79) 0.003

Systemic inflammatory responsemGPS (0/1/2) 2.12 (1.41–3.20) <0.001

Neutrophil-Lymphocyte Ratio (<5/>5) 2.27 (0.99–5.19) 0.052

Local inflammatory cell infiltrateManual Klintrup–Makinen grade (0/1/2/3) 0.41 (0.26–0.65) <0.001Manual Klintrup–Makinen category (weak/strong) 9.48 (2.27–39.58) 0.002Automated Klintrup–Makinen grade (quartiles – 0/1/2/3) 0.59 (0.42–0.82) 0.002Automated Klintrup–Makinen category (weak/strong) 0.41 (0.20–0.82) 0.012

Fig. 2a. Kaplan–Meier survival curve demonstrating the relationshipbetween Manual Klintrup–Makinen grade and cancer-specificsurvival.

Fig. 2b. Kaplan–Meier survival curve demonstrating the relationshipbetween Manual Klintrup–Makinen category (weak/strong) andcancer-specific survival.

R. Forrest et al. / European Journal of Cancer 50 (2014) 544–552 551

Importantly, in terms of potential clinical utility, theautomated assessment of the local inflammatory infil-trate using the K–M criteria is simple, practical andquick to perform. In addition, further software develop-ments including refinements to image annotation mayenhance clinical utility of the methodology further. Suchdevelopments, for example the placeable grid, may

facilitate an even more practical and time effectiveassessment of the local inflammatory response that canbe used widely in the prognostic assessment of patientswith colorectal cancer.

Given the importance of the host local immuneresponse in colorectal cancer progression, there havebeen calls to incorporate an assessment of this into

Fig. 3a. Kaplan–Meier survival curve demonstrating the relationshipbetween Automated Klintrup–Makinen grade and cancer-specificsurvival.

Fig. 3b. Kaplan–Meier survival curve demonstrating the relationshipbetween Automated Klintrup–Makinen category (weak/strong) andcancer-specific survival.

552 R. Forrest et al. / European Journal of Cancer 50 (2014) 544–552

routine clinical practice. Indeed, assessment of the localinflammatory cell response has the potential to predictoutcome and identify patients at risk of recurrence andfor whom adjuvant treatment should be considered [12].

In summary, the results of the present study demon-strate that automated assessment appears to effectivelyrecapitulate the clinical value of visual assessment ofthe local inflammatory cell infiltrate at the invasive mar-gin of colorectal tumours.

In addition, the present study demonstrates it is pos-sible to obtain an objective assessment of tumourinflammatory infiltrates using routinely stained H&Esections and that an automated, computer-based scoring

method is therefore a workable and cost-effectiveapproach to the clinical assessment of local immune cellinfiltrates in patients with colorectal cancer.

Conflict of interest statement

None declared.

Appendix A. Supplementary data

Supplementary data associated with this article canbe found, in the online version, at http://dx.doi.org/10.1016/j.ejca.2013.11.003.

References

[1] Oliphant R, Nicholson G, Horgan P, Molloy R, McMillan D,Morrison D. Deprivation and colorectal cancer surgery: longer-term survival inequalities are due to differential postoperativemortality between socioeconomic groups. Ann Surg Oncol2013:1–8.

[2] Klintrup K, Makinen JM, Kauppila S, Vare PO, Melkko J,Tuominen H, et al. Inflammation and prognosis in colorectalcancer. Eur J Cancer 2005;41(17):2645–54 [Epub 2005/10/22].

[3] Galon J, Costes A, Sanchez-Cabo F, Kirilovsky A, Mlecnik B,Lagorce-Pages C, et al. Type, density, and location of immunecells within human colorectal tumors predict clinical outcome.Science 2006;313(5795):1960–4.

[4] Roxburgh CS, McMillan DC. The role of the in situ localinflammatory response in predicting recurrence and survival inpatients with primary operable colorectal cancer. Cancer TreatRev 2012;38(5):451–66 [Epub 2011/09/29].

[5] Taylor CR, Levenson RM. Quantification of immunohistochem-istry – issues concerning methods, utility and semiquantitativeassessment II. Histopathology 2006;49(4):411–24 [Epub 2006/09/19].

[6] Gokhale S, Rosen D, Sneige N, Diaz LK, Resetkova E, Sahin A,et al. Assessment of two automated imaging systems in evaluatingestrogen receptor status in breast carcinoma. Appl Immunohis-tochem Mol Morphol 2007;15(4):451–5 [Epub 2007/12/20].

[7] Allard MA, Bachet JB, Beauchet A, Julie C, Malafosse R, PennaC, et al. Linear quantification of lymphoid infiltration of thetumor margin: a reproducible method, developed with colorectalcancer tissues, for assessing a highly variable prognostic factor.Diagn Pathol 2012;7:156 [Epub 2012/11/15].

[8] Vayrynen JP, Vornanen JO, Sajanti S, Bohm JP, Tuomisto A,Makinen MJ. An improved image analysis method for cellcounting lends credibility to the prognostic significance of T cellsin colorectal cancer. Virchows Arch 2012;460(5):455–65 [Epub2012/04/25].

[9] AJCC. Colon and rectum cancer staging. AJCC cancer stagingmanual. 7th ed.; 2010.

[10] Rizzardi AE, Johnson AT, Vogel RI, Pambuccian SE, HenriksenJ, Skubitz AP, et al. Quantitative comparison of immunohisto-chemical staining measured by digital image analysis versuspathologist visual scoring. Diagn Pathol 2012;7:42 [Epub 2012/04/21].

[11] Halama N, Zoernig I, Spille A, Westphal K, Schirmacher P,Jaeger D, et al. Estimation of immune cell densities in immune cellconglomerates: an approach for high-throughput quantification.PLoS One 2009;4(11):e7847 [Epub 2009/11/20].

[12] Galon J, Pages F, Marincola FM, Thurin M, Trinchieri G, FoxBA, et al. The immune score as a new possible approach for theclassification of cancer. J Transl Med 2012;10:1 [Epub 2012/01/05].