automated classification of breast pathology using local

Dartmouth College Dartmouth College

Dartmouth Digital Commons Dartmouth Digital Commons

Dartmouth Scholarship Faculty Work

11-1-2010

Automated Classification of Breast Pathology using Local Automated Classification of Breast Pathology using Local

Measures of Broadband Reflectance Measures of Broadband Reflectance

Ashley M Laughney Dartmouth College

Venkataramanan Krishnaswamy Dartmouth College

Pilar Beatriz Garcia-Allende University de Cantabria

Olga M Conde University de Cantabria

Wendy A Wells Dartmouth College

See next page for additional authors

Follow this and additional works at httpsdigitalcommonsdartmouthedufacoa

Part of the Engineering Commons and the Medicine and Health Sciences Commons

Dartmouth Digital Commons Citation Dartmouth Digital Commons Citation Laughney Ashley M Krishnaswamy Venkataramanan Garcia-Allende Pilar Beatriz Conde Olga M Wells Wendy A Paulsen Keith D and Pogue Brian W Automated Classification of Breast Pathology using Local Measures of Broadband Reflectance (2010) Dartmouth Scholarship 3711 httpsdigitalcommonsdartmouthedufacoa3711

This Article is brought to you for free and open access by the Faculty Work at Dartmouth Digital Commons It has been accepted for inclusion in Dartmouth Scholarship by an authorized administrator of Dartmouth Digital Commons For more information please contact dartmouthdigitalcommonsgroupsdartmouthedu

Authors Authors Ashley M Laughney Venkataramanan Krishnaswamy Pilar Beatriz Garcia-Allende Olga M Conde Wendy A Wells Keith D Paulsen and Brian W Pogue

This article is available at Dartmouth Digital Commons httpsdigitalcommonsdartmouthedufacoa3711

Journal of Biomedical Optics 15(6) 066019 (NovemberDecember 2010)

Automated classification of breast pathology using localmeasures of broadband reflectance

Ashley M LaughneyVenkataramanan KrishnaswamyDartmouth CollegeThayer School of Engineering8000 Cummings HallHanover New Hampshire 03755

Pilar Beatriz Garcia-AllendeUniversity of CantabriaPhotonics Engineering GroupAvda de los Castros SNSantander 39005 Spain

andImperial College LondonDepartment of Surgery and CancerExhibition RoadLondon SW7 2AZ United Kingdom

Olga M CondeUniversity of CantabriaPhotonics Engineering GroupAvda de los Castros SNSantander 39005 Spain

Wendy A WellsDartmouth-Hitchcock Medical CenterDepartment of Pathology1 Medical Center DriveLebanon New Hampshire 03756

Keith D PaulsenBrian W PogueDartmouth CollegeThayer School of Engineering8000 Cummings HallHanover New Hampshire 03755

Abstract We demonstrate that morphological features pertinent toa tissuersquos pathology may be ascertained from localized measures ofbroadband reflectance with a mesoscopic resolution (100-μm lateralspot size) that permits scanning of an entire margin for residual diseaseThe technical aspects and optimization of a k-nearest neighbor classifierfor automated diagnosis of pathologies are presented and its efficacy isvalidated in 29 breast tissue specimens When discriminating betweenbenign and malignant pathologies a sensitivity and specificity of 91and 77 was achieved Furthermore detailed subtissue-type analysiswas performed to consider how diverse pathologies influence scatteringresponse and overall classification efficacy The increased sensitivity ofthis technique may render it useful to guide the surgeon or pathologistwhere to sample pathology for microscopic assessment Ccopy2010 Society ofPhoto-Optical Instrumentation Engineers [DOI 10111713516594]

Keywords vis-NIR spectroscopy backscattering tissues biomedical optics

Paper 10236RR received May 2 2010 revised manuscript received Sep 26 2010accepted for publication Oct 4 2010 published online Dec 30 2010

1 IntroductionBreast-conserving therapy (BCT) which includes local exci-sion and radiation treatment to the breast has been the standardof care for early invasive breast cancers (stages I and II) sinceFisher and Veronisi demonstrated that BCT is equally safe andeffective as mastectomy for most patients when surgical marginsare clear of residual disease1 However the number of patientswith residual disease at or near the cut edge of their primarysurgical specimen ranges from 5 to 82 with the majority ofstudies indicating positive margins in 20ndash40 of patients2 Thishigh variability indicates that surgical margin assessment lacksstandardization despite its importance to a patientrsquos long-termprognosis In a large retrospective analysis of frozen sections

Address all correspondence to Ashley M Laughney Dartmouth College ThayerSchool of Engineering 8000 Cummings Hall Hanover New Hampshire 03755Tel 802-343-8737 E-mail ashleylaughneydartmouthedu

by Rogers et al3 and Laucirica4 diagnostic errors related tointraoperative evaluation of involved margins were attributed tointerpretation (57) microscopic sampling (24) gross sam-pling (95) and lack of communication between pathologistand surgeon (95) Effective margin assessment evidently re-quires a wide-field imaging technology with resolution sensitiveto diagnostically differentiating volumes of tissue and automatedinterpretation of results To reduce sampling error a broadbandreflectance imaging system was designed to provide localizedmeasures of spectroscopic information at a resolution that per-mitted wide-field scanning of a surgical specimen5 To addressinterpretation errorscattering maps were associated with diag-nostically relevant pathologies using an automated k-nearestneighbor (k-NN) classifier6 The effective illumination volumeof the imaging system (100-μm lateral resolution) was specifi-cally chosen to optimize direct sampling of elastic scattering

1083-3668201015(6)06601912$2500 Ccopy 2010 SPIE

Journal of Biomedical Optics NovemberDecember 2010 Vol 15(6)066019-1

Laughney et al Automated classification of breast pathology using local measures of broadband reflectance

features which are highly sensitive to tissue architecturalchanges at the cellular and subcellular levels7 8 These scatteringfeatures are quite relevant to the tissue parameters a patholo-gist would address during microscopic assessment of a surgicalspecimen The ability of the k-NN classifier to accurately de-tect morphological variations in tissue was first demonstratedin a mouse pancreas model5 6 and here its ability to discrimi-nate pathologies was demonstrated in human breast tissues Themorphology of pancreas and breast tumors are fundamentallydifferent the former is often dominated by abundant scleroticmalignant stroma while the latter usually has a more prominentmalignant epithelial component

A positive surgical margin is a major predictor of localrecurrence9ndash12 consequently the standard of care at Dartmouth-Hitchcock Medical Center (DHMC) for more than seven yearshas been to reexcise a margin if invasive tumor is ldquoatrdquo thatmargin andor if ductal carcinoma in situ (DCIS) is lt01 cmfrom the margin (WAW) The poor prognosis associated withpositive margins necessitates this standard of care despite thewell-documented psychological effects and suboptimal cos-metic results following repeated surgery13 14 Current modali-ties of intraoperative margin assessment include gross examina-tion of the specimen frozen section analysis (FSA) and touchpreparation cytology4 15 16 Each has its limitations gross eval-uation of the specimen does not reflect the microscopic mar-gin status15 FSA often results in severe histological artifacts(particularly in the breast because of its high adipose content)and requires substantial processing time (sim30 min requiringimmediate access to a cryostat and pathologist)2 16 and touchpreparation cytology is undesirable because of its inability tolocalize a ldquopositiverdquo measure15 New techniques for margin de-lineation must be simple rapid and nondisruptive to surgicalworkflow Intra-operative assessment of sentinel lymph node(s)could also have significant clinical impact during BCT as re-cently demonstrated by Austwick17 Principle components ex-tracted from elastic scattering spectra were classified accordingto pathology and the need for axial lymph node dissection wasdetermined with a sensitivity and specificity of 69 and 96respectively17

Spectroscopy methods have been studied extensively as di-agnostic tools for identifying residual disease in involved surgi-cal margins particularly fluorescence diffuse reflectance andRaman spectroscopy have been used to provide biochemicalinformation and scattering spectroscopy has been used to pro-vide ultrastructural information about a tissue18ndash23 The effectiveillumination volume determines the transport model and conse-quently the type of information obtained because measures areaveraged over different tissue volumes likely in different phys-iological states Techniques that probe larger volumes of tis-sue and diffuse light rely on the assumption that ultrastructuralmalignant transformations provide disease-specific contrast involume-averaged measures but sensitivity of the measurementto specific features of the tissue is not yet clear Backmanet al7 24 and Perelman et al8 demonstrated that wavelength-dependent variations in the intensity of singly backscatteredphotons were sensitive to nuclear enlargement pleomorphismcrowding and hyperchromatism7 8 24 Using light-scatteringspectroscopy (LSS) they observed alterations in the size andnumber density of epithelia in Barrettrsquos esophagus the bladderoral cavity and colon25ndash28 Georgakoudi augmented LSS with

diffuse reflectance spectroscopy and fluorescence spectroscopy(trimodal spectroscopy) to discriminate between dysplasia orcancerous and normal oral tissues achieving a sensitivity andspecificity of 9628 Keller et al29 Breslin et al30 Palmeret al31 and Zhu et al32 combined diffuse reflectance withfluorescence spectroscopy to discriminate between benign andmalignant breast tissues achieving a sensitivity specificity of700917 and 7899 respectively More recently Kelleret al29 developed a Monte Carlo inverse model to generate chro-mophore and scattering maps from visiblendashnear-infrared diffusereflectance spectra When applied to 55 surgical margins a mul-tivariate predictive model discriminated between positive andnegative breast tissue margins up to 2 mm in depth with a sen-sitivity and specificity of 79 and 67 respectively33 AmelinkSterenborg Amelink et al and van Veen et al34ndash36 developeddifferential path-length spectroscopy to determine local opti-cal properties of breast tissue in vivo using a fiber-optic needleprobe with optimized source-detector separation The sensingdepth of the probe was specifically limited to 240 μm so thatthe photon pathlength was independent of wavelength allowingstraightforward interpretation of the measured signal in terms ofabsorption and scattering34ndash36

The confocal measurement geometry employed here was de-signed to minimize multiple scattering and absorption effectstherefore the analysis focuses only on the spectral response ofphotons experiencing few scattering events and their value toclassification The lateral resolution was fixed at 100 μm (ap-proximately one mean scattering pathlength in tissue) becausethis obstructed multiply scattered light537ndash39 A strong period-icity in wavelength related to the size of scattering centers wasnot observed in localized tissue spectra because of the tissuersquosbroad particle size distribution and because the system does notexclusively look at single scattering events However extractedscattering features in this mesoscopic volume demonstrate sensi-tivity to the tissue parameters a pathologist would address duringmicroscopic assessment of a surgical specimen Scattering mea-sures are more robust and localized as compared to biochemicalparameters which tend to be diffuse transient and significantlydifferent in vitro as compared to in vivo Additionally scatteringparameters have not been evaluated sufficiently in a wavebandthat avoids coupling with absorption A k-NN algorithm au-tomated classification of spectral measures according to tissuepathology and was employed to guide the pathologist or surgeonwhere to look when evaluating involved surgical margins Thek-NN classifier was selected because it is simple nonparamet-ric robust has real-time capabilities and good performance foroptimal values of k

2 Materials and Methods21 Localized Reflectance SpectroscopyFresh breast tissue specimens were imaged in a custom-builtmicrosampling reflectance spectral imaging system5 Thissystem employs a quasi-confocal illumination and detection(510ndash785 nm) to constrain the overlapping illumination anddetection spot sizes to within approximately one scatteringdistance in tissue (sim100 μm in the visible) A complete descrip-tion of this imaging system can be found in a previous study5

Sampling in this mesoscopic regime allows the use of simple



Fig 1 (a) Representative HampE sections of breast tissue illustrating morphological features characteristics of distinct diagnostic categories in the breastand (b) illustrative relationship between radiation transport length scales and the primary scattering centers in breast tissue

empirical models to describe the light transport (Fig 1) For theshort pathlengths involved and for typical values of absorptionand scattering in tissue the measured spectral response is pro-portional to the reduced scattering coefficient μsrsquo In regionswhere significant local absorption is encountered a Beerrsquoslawndashtype attenuation factor is used to correct for the effects ofabsorption537ndash39 Each tissue sample was mounted on a glassslide and surveyed across the illumination-detection path usinga motorized translational stage Each measurement took sim2 hand during scanning the sample was hydrated with a phosphatebuffer solution Trace background reflections from the opticalsystem were acquired (Rbkgrd) and subtracted from measuredspectra and the data were normalized to the instrumentalspectral response (RSpectralon) of the system using a Spectralonreference (Labsphere Inc North Sutton New Hampshire)as follows

R(λ) = Racquired(λ) minus Rbkgrd(λ)

RSpectralon(λ) minus Rbkgrd(λ) (1)

Referencing to Spectralon also permitted direct comparisonbetween tissue samples and provided a daily calibration

An empirical approximation to Mie theory was used to de-scribe the measured reflectance spectrum R(λ) from a volume-averaged region of tissue40 Additionally a Beerrsquos Law attenu-ation factor is required to correct for the presence of significantlocal absorption by hemoglobin

R(λ) = Aλminusb expminuslowast[HbT]StO2lowastεHbO2 (λ)+(1minusStO2)lowastεHb(λ) (2)

Parameters A and b are scattering amplitude and scatteringpower respectively Both depend on the size and number den-sity of scattering centers in the volume of interrogated tissuethereby reflecting variations in breast tissue morphology41ndash43

refers to the mean optical pathlength (dependent on the illu-mination and detection geometry) parameter [HbT ] is the totalhemoglobin concentration parameter StO2 is the oxygen sat-uration factor (ratio of oxygenated to total hemoglobin) and

εHbO2 and εHb refer to the molar extinction coefficients of thesetwo chromophores respectively (Oregon Medical Laser CenterDatabase44) Oxygenated and deoxygenated hemoglobin werethe dominant tissue chromophores encountered in the measuredwaveband Measured reflectance spectra were fit to this modelusing a nonlinear least-squares solver to obtain estimates ofscattering amplitude and scattering power relative to SpectralonA measure of average scatter irradiance Iavg was calculated byintegrating the reflectance spectrum over a waveband that avoidsthe hemoglobin absorption peaks (620785 nm) Scattering pa-rameters were then microscopically correlated to morphologicalfeatures identified by a pathologist (WAW) on hematoxylin andeosin (HampE) stained sections of the tissue cut in the exact samegeometry as imaged in situ

22 Breast Tissue Surgical SpecimensAll studies were completed based on a protocol approved by theCommittee for the Protection of Human Subjects InstitutionalReview Board at Dartmouth Fresh breast tissue was obtaineddirectly from the Department of Pathology at DHMC from pa-tients who had given informed consent to allow this use of theirtissue Samples were procured during conservative surgery orbreast-reduction surgery and only provided if there was tissue inexcess of that required to make a clinical diagnosis Tissue sam-ples were 1ndash2 cm2 with a thickness of about 3ndash5 mm Sampleswere imaged within 12 h of surgery and in the case of delaythe tissue was stored in a 4C refrigerator and hydrated with aphosphate buffer solution Immediately following imaging eachsample was placed in 10 formalin and processed for histology(paraffin embedded sectioned to 4 μm and stained with HampE)A total of 35 tissue specimens were imaged 6 were rejected fromanalysis due to a low signal-to-noise ratio andor poor histolog-ical processing (both a consequence of highly fatty tissue) Inthe remaining 29 tissue samples 48 regions of interest (ROI)



Table 1 Distribution of the sample population according to tissuetype and subtype

Tissue Type and Subtype No ROI No Spectra

Total Not Malignant 25 36979

Normal Epithelia and Stroma 21 31226

Benign Epithelia and Stroma 3 5220

Inflammation 1 533

Total Malignant 14 23220

DCIS 1 194

IDC 12 22547

ILC 1 479

Total Adipose 9 7021

Adipose 9 7021

Total ROI 48 67220

were identified by the pathologist these are summarized inTable 1 The pathologist identified seven tissue pathologies in thesamples which were classified more generally as not-malignantmalignant or adipose

Figure 2(a) illustrates coregistration between the white lightimage histology and images of scattering parameters for a tissuesample Figures 2(b) and 2(c) show box plots of the scatteringpower and the logarithm of the wavelength-integrated irradiancewith outliers removed (those of gt2 standard deviations from themean) for all tissue samples The scattering amplitude is notdisplayed because it follows the same trend as the scatteringpower per diagnostic category and a correlation is observedbetween the scattering power and the logarithm of the scatteringamplitude [Fig 2(d)]

Histopathology reveals that the three macroscopic scatteringcenters found in breast tissue are epithelia stroma and adiposeImmunohistochemistry shows that the percent distribution ofthese components varies with diagnosis and registration ofscattering maps with pathology illustrate how spectral responsechanges as a function of diagnosis This suggests the percentdistribution of stroma epithelia and adipocytes in the effectiveillumination volume influences scattering response Standardimmunohistochemistry techniques were used to assess the per-cent distribution of adipose stroma and epithelia per sampleFormalin-fixed and paraffin-embedded tissue sections were cutand mounted on OptiPlusTM Positive Charged Barrier slides(BioGenex San Ramon California) to test for anti-cytokeratins8 and 18 (clone 5D3 BioGenex San Ramon California) Inbrief slides were deparaffinized with xylene and hydratedthrough graded alcohols and epitope retrieval was carried outin a steamer under pressure Slides were then rinsed in distilledwater soaked in phosphate-buffered saline and immunostainedusing the BioGenex i6000TM Automated Staining System(San Ramon California) Diaminobenzidine was applied forvisualization and hematoxylin was used as a counterstain

Slides were then dehydrated through graded alcohols andcoverslips were applied Whole immunostained slides were dig-itally scanned and montaged using the Surveyor Rcopy AutomatedSpecimen Scanning (Objective Imaging Ltd CambridgeUnited Kingdom) automated stage control bundled softwareThe epithelia-to-stroma ratio was quantified using Image-ProPlus (Media Cybernetics Bethesda Maryland) image-analysissoftware The epithelial and stromal percentages were definedas the percent of CK5D3 positive or hematoxylin counter-stained tissue thresholded in pseudocolor in the diagnosticROI as compared to the total area of the tissue respectivelyFigure 3(a) shows fitted spectra sampled from normal benignand malignant tissues respectively Figure 3(b) illustrates howepithelia stroma and fat content vary between normal benignand malignant samples based on this analysis

23 k-NN ClassificationThe distribution of scattering parameters demonstrated subtlediscrimination between tissue subtypes but data were multi-parametric and overall classification was challenging A k-NNclassifier was employed for ready discrimination between tis-sue pathologies The k-NN classifier simultaneously interpretsmultiple scattering parameters for tissue characterization by as-signing an unclassified vector (herein referred to as the querypoint) to the majority diagnosis of its k-nearest vectors foundin the feature space The feature space for three scatter-relatedparameters (scattering amplitude scattering power and totalwavelength-integrated intensity) is depicted in Fig 2(d) as wellas a query point with unknown diagnosis All tissue pixels weredefined according to a vector in the three-dimensional (3-D)feature space and were assigned to the training set (populatedfeature space) or to the validation set (query points) Sampledistributions between training and validation sets were madeboth randomly and according to a leave-one-out analysis perpatient45 All training pixels were associated with a known di-agnosis according to the pathologistrsquos demarcation of ROI Thediagnosis of each query point was also determined by the pathol-ogist but remained unknown to the classifier in order to evaluateits performance

Additional feature extraction from the actual data set hasbeen shown to compensate for pixel-to-pixel variations and toimprove the overall performance of the classifier6 Therefore thefirst four statistical moments (mean standard deviation skew-ness and kurtosis) of each scattering parameter were estimatedin a real two-dimensional (2-D) spatial window centered abouteach pixel of interest These local statistical parameters wereconcatenated to the actual scatter parameters and parametricfeature space was expanded from three-dimensions to 15 di-mensions The behavior of the classifier was then studied as afunction of two independent variables the number of nearestneighbors k and the size of the spatial window used to computelocal statistics

Accuracy of the k-NN classifier fundamentally depends onthe metric used to compute distances between the query pointand all training pixels in the feature space To prevent somefeatures from being more strongly weighted than others datamust be transformed so that all parameters are of the sameorder of magnitude To standardize the dynamic range of thescattering parameters scattering amplitude and average scatterirradiance were log-transformed All scattering parameters



Fig 2 (a) Coregistration between the digital photograph histology and maps of scattering power amplitude and total-wavelength integratedintensity for a given tissue sample (b c) box plots of relative scatter power and log of the total-wavelength integrated intensity according to diagnosis(outliers gt 2 std not displayed) and (d) the 3-D feature space assembled with scattering parameters and employed by the k-NN classifier

were statistically normalized to zero mean and unit varianceand outliers were removed from the feature space (training set)according to their interquartile fractions Query points werenever marked as outliers because this information would not beknown a priori When no prior knowledge is available aboutthe probability density function of a particular class (diagnosis)

most distance-based classifiers employ a Euclidean distancemetric [Eq (3)] where M is the identity matrix The Euclideanmetric assumes all parameters defining a pixel are equallyimportant and independent of the others46

d(x y)2 = (y minus x)T M(y minus x) (3)



Fig 3 (a) The fitted spectra sampled from normal benign and malignant tissues respectively are shown (b) The distribution of stroma epitheliaand adipose are shown across the three diagnostic categories classified by immunohistochemistry for all tissue samples

However this metric is not ideal for this data set becauseevaluation of feature space reveals strong coupling betweenscattering amplitude and scattering power as evident in Fig 2(d)Statistical regularities were estimated from a large training setand the Mahalanobis distance metric was employed to accountfor correlations between parameters The Mahalanobis distancemetric is given in Eq (3) where M is the covariance matrix forthe scattering parameters In the special case where features areuncorrelated and variance is the same for all parameters theMahalanobis metric is equivalent to the Euclidean metric Asignificant improvement in the k-NN classifierrsquos performancewas observed using the Mahalanobis metric as compared tothe Euclidean metric for distance calculations with this dataset

24 Validation of the ClassifierIn order to optimize the independent variables associated withthe classifier a threefold cross-validation technique was ap-plied for discrimination between not-malignant malignant andadipose samples and for discrimination between all pathology

subtypes identified in Table 147 48 All data were randomly di-vided into three nonoverlapping sets with an equal number ofpixels per diagnostic category per set Two of these sets wereemployed as a training set (used to populate feature space)and the other was employed as a validation set (query points)to compute the classification error Error was taken to be thepercentage of misclassified pixels in the validation set where amisclassification means that the diagnosis assigned to a pixel bythe automated classifier does not match the diagnosis providedby the pathologist This procedure was repeated three times forall possible permutations of training and testing sets and thereported classification error is the average of these three execu-tions This threefold cross-validation was repeated for a varyingnumber of nearest neighbors k and a varying spatial windowsize for computation of local statistics

Additionally leave-one-out analysis was performed per pa-tient where ROI from one tissue sample populate the validationset and all other ROI pixels populate the feature space In thisvalidation procedure points are not equally distributed betweendiagnostic categories in either the training or testing sets Im-ages of the classification results were generated in HampE false



Fig 4 Evaluation of k-NN classification error as a function of the number of nearest neighbors k and as a function of the local window size usedto compute higher order statistics for discriminating between all pathologies and between not-malignant malignant and adipose pathologies

color for each tissue sample allowing one to evaluate whetherthe predicted diagnosis outside selected ROI makes sense in thecontext of the entire sample A mode filter was applied in asliding window (5times5 pixels) over the k-NN classified image toeliminate impulsive assignment noise The error and efficacy ofthe classifier was summarized for all tissue samples

Pixels corresponding to locations where reflectance spectracould not be reliably measured were tagged as masked pixelsand these were excluded from the training and validation setsduring all cross-validation procedures

3 Results31 Optimization of Independent Parameters in the

ClassifierFigure 4 depicts the behavior of the classification algorithmwhen discriminating between all pathologies (top lines on plot)and when discriminating more generally between not-malignantmalignant and adipose tissues (bottom lines) Classification isrepeated as a function of the spatial window used to computelocal statistics and the number of k-NNs Error is slightly higherwhen classifying all pathologies because the seven diagnosticcategories identified by the pathologist are not equally repre-sented with regard to the number of data points in feature spaceTable 1 indicates that there are significantly more points in thefeature space for the subtypes normal epithelia and stromainvasive ductal carcinoma (IDC) invasive lobular carcinoma(ILC) and adipose The more general classification scheme wasadopted to compensate for an uneven distribution of pixels perdiagnostic category The error asymptotically approaches sim2for a decreasing number of nearest neighbors k and increasingspatial window for local statistics for both levels of classificationHowever this accuracy is misleading and several features of thegraph indicate that the classifier is governed by statistics ratherthan the actual scattering parameters when the spatial window

for local statistics is gt4 pixels (gt400 μm) These indicatorsinclude the following

1 As the spatial window for local statistics increases in sizethe distributions used to compute local statistical parame-ters will become more normal and error should decreasehowever if the size of the window becomes too great pix-els from different diagnostic categories will be mixed anderror will increase Therefore classification error shoulddecrease with increasing spatial window size when thelocal window maintains separation between diagnosticcategories Because a consistent asymptotic decrease isobserved in classification error for all evaluated local win-dows for statistics the classifier is clearly governed by thestatistical parameters for a large local window

2 Classification accuracy is expected to increase with thenumber of nearest neighbors because this reduces the in-fluence of outliers rendering a diagnosis more probabilis-tic This is observed when statistics are computed in a localwindow of lt400 μm2 surrounding the pixel of interestFor a larger local window classification error increaseswith the number of nearest neighbors

On the basis of these observations a spatial window of400 μm2 was chosen to compute local statistical parametersand seven nearest neighbors were considered during automateddiagnosis of pixels for the leave-one-out patient analysis A 400μm2 window for computation of local statistics is physicallyreasonable because it approaches the mode of ROI dimension-ality while preserving spatial resolution These considerationswere made so that the raw scattering parameters not their statis-tics govern the classifierrsquos behavior For 7-NN and a 400-μm2

window for computation of local statistics Fig 4 approximates14 and 8 error during leave-one-out analysis when discrim-inating between all pathologies and not-malignant malignantand adipose tissue respectively



Table 2 Summary of the classification error and total efficacy of the k-NN classifier when discriminating between not-malignant malignant andadipose tissue and when discriminating between all pathologies Reported measures based on ability to discriminate given pathology from all otherdiagnostic categories evaluated

Classification

(Not-malignant Classification

Malignant Fat) (All Pathologies)

Classification Error

Median 875 168

Mean 130 253

Standard Deviation 137 250

Interquartile range 155 255

[min max] [0 535] [215 953]

Total Efficacy Not-malignant Malignant Fat Normal Benign DCIS IDC ILC Inflam Fat

Accuracy 086 086 098 074 085 100 086 099 099 098

Sensitivity 090 077 087 074 009 000 077 000 000 087

Specificity 082 090 099 074 091 100 090 100 100 099

Negative Predictive Value 087 088 099 077 092 100 089 099 099 099

Positive Predictive Value 086 081 095 071 008 000 079 000 000 095

32 Discrimination between Tissue Pathologiesper Sample

Table 2 summarizes the classification efficacy and classificationerror observed when performing leave-one-out validation forall tissue samples The median classification error is approxi-mately 17 and 9 when discriminating between all pathologiesand not-malignant malignant and adipose tissue respectivelyThis is quite close to our performance estimates in Section 31When classifying all pathologies low sensitivity is observedfor those classes under-represented in sample space (benignDCIS IDC ILC and inflammation) In any case the classifierhas clinical application because normal epithelia and stromaIDC and adipose are the most frequently encountered tissuesduring conservative breast surgery The classifierrsquos sensitivity tonot-malignant malignant and adipose pathologies is 090 077and 087 respectively Sensitivity is lower in malignant samplesbecause its sample population is characterized by greater het-erogeneity Although DCIS IDC and ILC are all consideredmalignant morphologically and biologically they are quite dis-tinct Specificity of the classifier for not-malignant malignantand adipose pathologies is 082 090 and 099 respectivelySpecificity is lowest in normal tissues because these are charac-terized by mixed fibroglandular and adipose content Epithelialproliferation in malignant tissues was observed to crowd outadipocytes in this study In a reflectance geometry scatteringfrom adipocytes results in a very low (noisy) signal The nega-tive and positive predictive values for each diagnostic categoryare also reported These refer to the number of patients withnegative and positive results (respectively) who are correctly di-agnosed For surgical margin applications the surgeon is most

interested in a high negative predictive value ensuring hisherdiagnosis of normal or malignant is an accurate one The nega-tive predictive values for not-malignant and malignant patholo-gies are 87 and 88 respectively Although less essential highpositive predictive values prevent any unnecessary reexcisionsduring surgery

The confusion matrices in Table 3 illustrate the distribu-tion of misclassified pixels across diagnostic categories whenperforming leave-one-out analysis for two levels of diagnosticdiscrimination This is important to consider because cost to thepatient for misclassifying a normal pixel as benign is less thanthe cost to the patient for misclassifying a malignant pixel asnormal

Figure 5 illustrates classification of six representative tis-sue samples The first column contains a digital photographof each tissue sample taken immediately after spectral imag-ing this is the surgeonrsquos perspective Fibroglandular tissue iswhite adipose is yellow-orange and higher concentrations ofhemoglobin are red Histological sections were coregistered tothe scattering and white light images and are displayed in col-umn 2 Hematoxylin has a deep blue-purple color and stainsnucleic acids which are primarily located in the cell nucleiEosin is pink and stains proteins nonspecifically mainly thecytoplasm and stroma have varying degrees of pink stainingFat is not preserved during histological processing thus thisbecomes empty space on the slide Column 3 illustrates theROI identified by the pathologist and they are colored ac-cording to their true diagnostic category while column 4 con-tains images of the automated diagnosis provided by the k-NNclassifier



Table 3 (a) A confusion matrix under leave-one-out validation that represents trends in misclassification and accurate identification of benign andmalignant pathologies (b) A confusion matrix under leave-one-out validation that represents trends in misclassification and accurate identificationof all pathologies identified by WAW The matrices list the percentage of pixels classified correctly along the diagonal (gray) and incorrectly off ofthe diagonal Clear misclassifications are highlighted in bold

Fig 5 Classification of six representative tissue samples Each row corresponds to a different tissue sample and the following four images areco-registered (a) a white light image of the tissue (b) HampE stained section of tissue (c) true diagnosis of ROI identified by the pathologist (WAW)and (d) classification results when discriminating between not-malignant malignant and adipose tissues



We recognize that the surgeon is most interested in a diagno-sis of either benign or malignant therefore the k-NN classifierwas executed with this binary level of discrimination Whenseparating benign and malignant pathologies the sensitivityand specificity was 91 and 77 respectively and the systemachieved positive and negative predictive values of 88 and 81respectively

4 DiscussionThis study demonstrates that morphological features pertinentto a tissuersquos diagnosis may be ascertained from confocal de-tection of broadband reflectance with a mesoscopic resolutionthat permits scanning of an entire margin for residual diseaseThe technical aspects and optimization of a k-NN classifier forautomated diagnosis of pathologies is presented and validatedin 29 specimens of breast tissue The classifierrsquos discriminat-ing capabilities improved with the inclusion of local statis-tics likely accounting for microscopic tissue heterogeneitiesInitially discrimination between all pathologies identified byWAW was attempted however inadequate sampling of uncom-mon pathologies rendered their classification less robust Giventhe sample population discrimination between not-malignantmalignant and adipose pathologies was most intuitive particu-larly because these diagnostic categories correspond to the threemacroscopic scattering centers in breast tissue (stroma epithe-lia and adipocytes) Negative predict values of 87 88 and 99were achieved for not-malignant malignant and adipose tis-sues respectively In the same order their positive predictivevalues were 86 81 and 95 The classifier was most sensi-tive to not-malignant and adipose tissues because the malignantpopulation was pathologically very diverse including samplesof ILC IDC and DCIS Specificity was lowest in not-malignantsamples because of their mixed fibroglandular and adipose con-tent In the context of conservative surgery the goal of treat-ment is to maximize removal of malignant tissues while min-imizing damage to healthy viable tissue When discriminat-ing between benign and malignant tissues only a sensitivityof 91 and a specificity of 77 was achieved This sensitiv-ity is significantly higher than those reported for frozen sectionanalysis (which is not practical for lumpectomy margins be-cause of the problems associated with freezing and cutting adi-pose tissues) and diffuse reflectance spectroscopy although itsspecificity is lower23 49 50 Even though overall efficacy of theclassifier exceeds or is comparable to other intraoperative as-sessment techniques integration of spectroscopy methods intothe surgical suite will require a better negative predictive valueensuring that the surgeonrsquos diagnosis of benign or malignantis an accurate one Additionally higher positive predictive val-ues would prevent unnecessary tissue removal during surgeryEfficacy of the proposed technique may be enhanced with betteror more data and further optimization of the classifier itself

This is a very purposeful application of spectroscopy be-cause it uses optical measures as defined according to pathologycriteria and may inform clinical practice during surgery Whilemechanically scanning the sample was time intensive and there-fore not suitable for clinical translation A second-generationsystem is under development that employs a scanning-beamarchitecture to image tissue fields up to 15times15 cm2 within 2ndash3 min The combination of tissue spectroscopy with automated

classification may provide information about tissue morphol-ogy intraoperatively so that if residual tumor is present at oneor more margins then the surgeon (before closing) could be ad-vised to remove more tissue immediatelymdashrather than at a laterreexcision Additionally wide-field optical maps generated bythe system may guide sampling of pathology by highlight sus-picious lesions for more detailed investigation Particularly thegreatest potential clinical impact could be intraoperative detec-tion of DCIS which has a very important status in breast pathol-ogy because its presence is a powerful predictor for local dis-ease reoccurrence and its identification during surgery influencestreatment options51ndash54 DCIS is a nonpalpable lesion unidenti-fiable during surgery consequently characterization of its spec-tral response could dramatically improve patient care Finallyunderstanding the relationship between the optical and biologi-cal properties of a tissue will ultimately improve the diagnosticutility of optical techniques permitting optimization of the mea-surement procedure and signal analysis for enhanced sensitivityto differentiating features The technique remains to be tested in-traoperatively Future clinical studies will reveal how the systemmay enhance existing surgery and pathologic procedures

One lesson learned over the course of this study was that re-flectance measures from tissues with high adipose content werecharacteristically low (and noisy) Our proposed technologicalsolution is automation of a variable integration time permit-ting longer data collection at points with a low signal until aminimum number of photons are detected Scrutiny of classi-fied tissue samples revealed that the greatest number of maskedpixels (a consequence of poor data fit to the empirical model)were found in regions of high adipose content and misclassi-fication frequently occurred in regions of mixed fibroglandu-lar and adipose content Garcia-Allende et al6 achieved nearlyperfect sensitivity and specificity when classifying tissue sub-types in pancreas tumors mainly because pancreas tissue ishighly stromal and almost completely lacking adipose contentAdditionally note that trends in scattering parameters are in-verted regarding diagnosis in the breast as compared to the pan-creas This is because malignancy is essentially stromal in thepancreas and epithelial in the breast and the percent distributionof stroma epithelia and adipocytes in the interrogation volumefundamentally influences the scattered signal there

Preliminary analysis of parameter covariance reveals a corre-lation between the logarithm of the relative scattering amplitudeand scattering power This correlation was also observed in aprevious study analyzing the scattering response of pancreastissue56 suggesting that the discriminating ability of the rela-tive scattering amplitude and power may be the same In thiswork they were treated independently during classification anda Mahalanobis distance metric was employed so that classifierefficacy was not influenced by their likeness in value Thebiological underpinnings of this correlation are not yet clearand beyond the scope of this paper However extracting in-dependent information from the relative scattering amplitudeand power will be an important focus of controlled phantomexperimentation

To expand on the variety of data collected a new scan-ning spectroscopy imaging system is being developed to allowbroadband spectral imaging of breast pathology in a wave-band (400ndash700 nm) that is sensitive to both tissue morphol-ogy and biochemical composition Particularly within this



waveband one may determine the concentration of oxygenatedand deoxygenated hemoglobin beta carotene and blood break-down products in a tissue while simultaneously extracting scat-tering features in a region with minimal absorption (gt620 nm)Addition of other optical parameters to the classifier is extremelysimple only involving an expansion of feature space (update thevector describing each pixel to N dimensions where N is thenumber of parameters) It would be particularly useful to gen-erate a comprehensive dataset with parameters describing allpossible endogenous lightndashtissue interactions (scattering ab-sorption fluorescence) so that the most diagnostically discrim-inating and robust parameters could be identified and optimizedduring data collection Note that beta-carotene is a member ofthe carotenoids and gives fat its highly pigmented yellow colorwe hope that its absorption spectra will improve classification oftissues with high adipose content Finally to enhance the speci-ficity of spectroscopic diagnosis one could envision the use ofRaman spectroscopy in suspicious regions identified by broad-band reflectance Nucleic acids proteins and lipids have distinctRaman features that provide highly specific structural and envi-ronmental information its primary limitation has always beensensitivity18

Improvement of the classifierrsquos performance may fundamen-tally be achieved with greater sampling As the number of datapoints in feature space increases so does the accuracy of theclassifier55 Particularly the classifierrsquos sensitivity to malignantpathologies has the most to gain and could be improved withequal representation of IDC ILC and DCIS in feature spaceAs feature space expands so will computational costs Ratherthan calculating the distance between each query point and everypoint in feature space a k-dimensional tree may be employed tooptimize the search algorithm56 The classifier was trained withROI obviously belonging to a diagnosis normal and malignantpathologies were identified by cellular features in fibroglandularregions and adipose ROI were measured far from any fibroglan-dular tissue Perhaps the performance of the classifier wouldimprove if an additional level of classification was employed sothat each diagnostic category was also labeled according to thesubtypes ldquofibroglandularrdquo or ldquofattyfibroglandularrdquo

5 ConclusionsIn this contribution we validate and optimize the ability of ak-NN classifier to automatically detect breast tissue patholo-gies based on direct sampling of elastic scattering features Thesampling volume was specifically chosen to be sensitive to archi-tectural changes addressed by a pathologist during microscopicassessment of a surgical specimen while also permitting itswide-field scanning Performance of the classifier was assessedin 29 breast tissue specimens and when discriminating betweenbenign and malignant pathologies a sensitivity and specificityof 91 and 77 was achieved Furthermore detailed subtissueanalysis was performed to consider how diverse pathologies in-fluence scattering response and overall classification efficacyClassification efficacy was influenced by the number of samplesused in training the level of discrimination used in separatingtissue types as well as morphological features inherent to eachpathology The increased sensitivity of this technique may ren-der it useful to guide the surgeon where to look when evaluatingan involved surgical margin or the pathologist where to sample

a gross specimen for microscopic assessment We expect thatspecificity and overall classification efficacy will improve withthe inclusion of additional optical parameters such as absorp-tion and fluorescence because these are responsive to the tissuersquosbiochemical state

AcknowledgmentsThis work was funded by National Institute of Health Grant NoPO1 CA80139 and the DoD CDMRP Grant No BC093811The authors also acknowledge the technical expertise of MarySchwab and Rebecca OrsquoMeara in the Department of Pathology atDHMC for their help with immunohistochemistry and histology

References1 F Fitzal O Riedl and R Jakesz ldquoRecent developments in breast-

conserving surgery for breast cancer patientsrdquo Langenbeckrsquos Archivesof Surgery 394(4) 591-609 (2009)

2 R Pleijhuis M Graafland J de Vries J Bart J de Jong and Gvan Dam ldquoObtaining Adequate Surgical Margins in Breast-ConservingTherapy for Patients with Early-Stage Breast Cancer Current Modali-ties and Future Directionsrdquo Annals of Surgical Oncology 16(10) 2717-2730 (2009)

3 C Rogers E C Klatt and P Chandrasoma ldquoAccuracy of Frozen-Section Diagnosis in a Teaching Hospitalrdquo Archives of Pathology ampLaboratory Medicine 111(6) 514-517 (1987)

4 R Laucirica ldquoIntraoperative assessment of the breast - Guidelinesand potential pitfallsrdquo Archives of Pathology amp Laboratory Medicine129(12) 1565-1574 (2005)

5 V Krishnaswamy P J Hoopes K S Samkoe J A OrsquoHara THasan and B W Pogue ldquoQuantitative imaging of scattering changesassociated with epithelial proliferation necrosis and fibrosis in tumorsusing microsampling reflectance spectroscopyrdquo J Biomed opt 14(1)014004 (2009)

6 P B Garcia-Allende V Krishnaswamy P J Hoopes K S SamkoeO M Conde and B W Pogue ldquoAutomated identification of tumormicroscopic morphology based on macroscopically measured scattersignaturesrdquo J Biomed opt 14(3) 034034 (2009)

7 V Backman et al Measuring cellular structure at submicrometerscale with light scattering spectroscopyrdquo IEEE Journal of SelectedTopics in Quantum Electronics 7(6) 887-893 (2001)

8 L T Perelman et al Observation of periodic fine structure in re-flectance from biological tissue A new technique for measuring nuclearsize distributionrdquo Physical Review Letters 80(3) 627-630 (1998)

9 C D Scopa P Aroukatos A C Tsamandas and C Aletra ldquoEvalua-tion of Margin Status in Lumpectomy Specimens and Residual BreastCarcinomardquo The Breast Journal 12(2) 150-153 (2006)

10 M C Smitt K W Nowels M J Zdeblick S Jeffrey R W Carlson FE Stockdale and D R Goffinet ldquoThe Importance of the LumpectomySurgical Margin Status in Long-Term Results of Breast-ConservationrdquoCancer 76(2) 259-267 (1995)

11 B Spivack M M Khanna L Tafra G Juillard and A E Giu-liano ldquoMargin Status and Local Recurrence after Breast-ConservingSurgeryrdquo Arch Surg-Chicago 129(9) 952-956 (1994)

12 S J Schnitt J R Harris and B L Smith ldquoDeveloping a prognosticindex for ductal carcinoma in situ of the breast - Are we there yetrdquoCancer 77(11) 2189-2192 (1996)

13 D E Wazer T Dipetrillo R Schmidtullrich L Weld T J Smith DJ Marchant and N J Robert ldquoFactors Influencing Cosmetic Outcomeand Complication Risk after Conservative Surgery and Radiotherapy forEarly-Stage Breast-Carcinomardquo J Clin Oncol 10(3) 356-363 (1992)

14 I A Olivotto M A Rose R T Osteen S Love B Cady B SilverA Recht and J R Harris ldquoLate Cosmetic Outcome after ConservativeSurgery and Radiotherapy - Analysis of Causes of Cosmetic FailurerdquoInt J Radiat Oncol 17(4) 747-753 (1989)

15 G C Balch S K Mithani J F Simpson and M C Kelley ldquoAc-curacy of Intraoperative Gross Examination of Surgical Margin Statusin Women Undergoing Partial Mastectomy for Breast MalignancyrdquoAmerican Surgeon 71(1) 22-28 (2005)



16 M R Wick and S E Mills ldquoEvaluation of surgical margins in anatomicpathology Technical conceptual and clinical considerationsrdquo SeminDiagn Pathol 19(4) 207-218 (2002)

17 M R Austwick et al Scanning elastic scattering spectroscopy detectsmetastatic breast cancer in sentinel lymph nodesrdquo J Biomed Opt 15(4)047001 (2010)

18 A Mahadevan-Jansen M F Mitchell N Ramanujam A Malpica SThomsen U Utzinger and R Richards-Kortum ldquoNear-infrared Ramanspectroscopy for in vitro detection of cervical precancersrdquo PhotochemPhotobiol 68(1) 123-132 (1998)

19 C J Frank R L McCreery and D C B Redd ldquoRaman-spectroscopyof normal and diseased human breast tissuesrdquo Anal Chem 67(5) 777-783 (1995)

20 I J Bigio et al Diagnosis of breast cancer using elastic-scatteringspectroscopy preliminary clinical resultsrdquo J Biomed Opt 5(2) 221-228 (2000)

21 I J Bigio and J R Mourant ldquoUltraviolet and visible spectroscopiesfor tissue diagnostics Fluorescence spectroscopy and elastic-scatteringspectroscopyrdquo Phys Med Biol 42(5) 803-814 (1997)

22 A C Lee C D O Pickard M R S Keshtgar G M BriggsM Falzon S Lakhani I Bigio and S G Bown ldquoElastic scatteringspectroscopy for the diagnosis of breast cancerrdquo Brit J Surg 89 74-74(2002)

23 T M Breslin G Palmer C Fang K W Gilchrist F Xu andN Ramanujam ldquoAuto fluorescence and diffuse reflectance propertiesof malignant and benign breast tissuesrdquo Annals of Surgical Oncology10(1) S16-S16 (2003)

24 V Backman et al Detection of preinvasive cancer cellsrdquo Nature406(6791) 35-36 (2000)

25 I Georgakoudi ldquoThe color of cancerrdquo J Lumin 119 75-83(2006)

26 I Georgakoudi et al Fluorescence reflectance and light-scatteringspectroscopy for evaluating dysplasia in patients with Barrettrsquos esoph-agusrdquo Gastroenterology 120(7) 1620-1629 (2001)

27 I Georgakoudi E E Sheets M G Muller V Backman C P CrumK Badizadegan R R Dasari and M S Feld ldquoTrimodal spectroscopyfor the detection and characterization of cervical precancers in vivordquoAm J Obstet Gynecol 186(3) 374-382 (2002)

28 M G Muller et al Spectroscopic detection and evaluation of mor-phologic and biochemical changes in early human oral carcinomardquoCancer 97(7) 1681-1692 (2003)

29 M D Keller S K Majumder M C Kelley I M Meszoely F IBoulos G M Olivares and A Mahadevan-Jansen ldquoAutofluorescenceand diffuse reflectance spectroscopy and spectral imaging for breastsurgical margin analysisrdquo Lasers in Surgery and Medicine 42(1) 15-23(2010)

30 T M Breslin F S Xu G M Palmer C F Zhu K W Gilchrist andN Ramanujam ldquoAutofluorescence and diffuse reflectance propertiesof malignant and benign breast tissuesrdquo Annals of Surgical Oncology11(1) 65-70 (2004)

31 G M Palmer P J Keely T M Breslin K W Gilchrist andN Ramanujam ldquoEx vivo diagnosis of breast cancer using fluorescenceand reflectance spectroscopyrdquo Lasers in Surgery and Medicine NationalMeeting (Atlanta GA) 76-76 (2002)

32 C Zhu G M Palmer T M Breslin J Harter and N RamanujamldquoDiagnosis of breast cancer using fluorescence and diffuse reflectancespectroscopy a Monte-Carlo-model-based approachrdquo J Biomed Opt13(3) 034015 (2008)

33 L G Wilke et al Rapid noninvasive optical imaging of tissuecomposition in breast tumor marginsrdquo Am J Surg 198(4) 566-574(2009)

34 A Amelink and H Sterenborg ldquoMeasurement of the local opticalproperties of turbid media by differential path-length spectroscopyrdquoApplied Optics 43(15) 3048-3054 (2004)

35 A Amelink H Sterenborg M P L Bard and S A BurgersIn vivo measurement of the local optical properties of tissue by useof differential path-length spectroscopyrdquo Optics Letters 29(10) 1087-1089 (2004)

36 R L P van Veen A Amelink M Menke-Pluymers C Van Der Poland H J C M Sterenborg ldquoOptical biopsy of breast tissue usingdifferential path-length spectroscopyrdquo Phys Med Biol 50(11) 2573-2581 (2005)

37 B W Pogue and T Hasan ldquoFluorophore quantitation in tissue-simulating media with confocal detectionrdquo IEEE Journal of SelectedTopics in Quantum Electronics 2(4) 959-964 (1996)

38 B W Pogue and G Burke ldquoFiber-optic bundle design for quantitativefluorescence measurement from tissuerdquo Applied Optics 37(31) 7429-7436 (1998)

39 J R Mourant J P Freyer A H Hielscher A A Eick D Shenand T M Johnson ldquoMechanisms of light scattering from biologicalcells relevant to noninvasive optical-tissue diagnosticsrdquo Applied Optics37(16) 3586-3593 (1998)

40 H J Vanstaveren C J M Moes J Vanmarle S A Prahl andM J C Vangemert ldquoLight-Scattering in Intralipid-10-Percent in theWavelength Range of 400-1100 Nmrdquo Applied Optics 30(31) 4507-4514 (1991)

41 M Bartek X Wang W Wells K D Paulsen and B W Pogue ldquoEsti-mation of subcellular particle size histograms with electron microscopyfor prediction of optical scattering in breast tissuerdquo J Biomed Opt11(6) 064007 (2006)

42 X Wang et al Image reconstruction of effective Mie scatteringparameters of breast tissue in vivo with near-infrared tomographyrdquo JBiomed Opt 11(4) 041106 (2006)

43 X Wang B W Pogue S D Jiang X M Song K D PaulsenC Kogel S P Poplack and W A Wells ldquoApproximation of Miescattering parameters in near-infrared tomography of normal breasttissue in vivordquo J Biomed Opt 10(5) 051704 (2005)

44 S L Jacques Prahl S ldquoOptical Properties Spectrardquo (Oregon MedicalLaser Center Portland OR 2010) httpomlcogiedu

45 O Eytan B A Sela and A Katzir ldquoFiber-optic evanescent-wave spec-troscopy and neural networks application to chemical blood analysisrdquoApplied Optics 39(19) 3357-3360 (2000)

46 D A Burns Ciurczak EW ldquoQualitative Discriminant AnalysisrdquoHandbook of Near-Infrared Analysis (Chpt 15) 3rd ed (CRC Press2008)

47 C Goutte ldquoNote on free lunches and cross-validationrdquo Neural Comput9(6) 1245-1249 (1997)

48 H Y Zhu and R Rohwer ldquoNo free lunch for cross-validationrdquo NeuralComput 8(7) 1421-1426 (1996)

49 N Cabioglu et al Role for Intraoperative Margin Assessment inPatients Undergoing Breast-Conserving Surgeryrdquo Annals of SurgicalOncology 14(4) 1458-1471 (2007)

50 G M Palmer C F Zhu T M Breslin F S Xu K W Gilchrist andN Ramanujam ldquoComparison of multiexcitation fluorescence and dif-fuse reflectance spectroscopy for the diagnosis of breast cancer (March2003)rdquo IEEE T Bio-Med Eng 50(11) 1233-1242 (2003)

51 N Q Mirza et al Ductal carcinoma-in-situ Long-term results ofbreast-conserving therapyrdquo Annals of Surgical Oncology 7(9) 656-664(2000)

52 E B Claus S Petruzella E Matloff and D Carter Prevalenceof BRCA1 and BRCA2 mutations in women diagnosed with ductalcarcinoma in siturdquo JAMA-Journal of the American Medical Association293(8) 964-969 (2005)

53 N S Goldstein L Kestin and F Vicini ldquoIntraductal carcinoma of thebreast - Pathologic features associated with local recurrence in patientstreated with breast-conserving therapyrdquo American Journal of SurgicalPathology 24(8) 1058-1067 (2000)

54 M R Bani M P Lux K Heusinger E Wenkel A Magener RSchulz-Wendtland M W Beckmann and P A Fasching ldquoFactorscorrelating with reexcision after breast-conserving therapyrdquo EuropeanJounral of Surgical Oncology 35(1) 32-37 (2009)

55 P B Garcia-Allende et al ldquoAutomated identification of tumor micro-scopic morphology based on microscopically measured scatter signa-turesrdquo J Biomed Opt 14(3) 034034 (2009)

56 K Fukunaga Introduction to Statistical Pattern Recognition (Aca-demic Press New York 1972)


Automated Classification of Breast Pathology using Local Measures of Broadband Reflectance

Dartmouth Digital Commons Citation

Authors

tmp1559313725pdfVHVBl

Authors Authors Ashley M Laughney Venkataramanan Krishnaswamy Pilar Beatriz Garcia-Allende Olga M Conde Wendy A Wells Keith D Paulsen and Brian W Pogue

This article is available at Dartmouth Digital Commons httpsdigitalcommonsdartmouthedufacoa3711















1083-3668201015(6)06601912$2500 Ccopy 2010 SPIE






























Inflammation 1 533


DCIS 1 194

IDC 12 22547

ILC 1 479


Adipose 9 7021

Total ROI 48 67220




































Classification




Median 875 168

Mean 130 253



[min max] [0 535] [215 953]


Accuracy 086 086 098 074 085 100 086 099 099 098

Sensitivity 090 077 087 074 009 000 077 000 000 087

Specificity 082 090 099 074 091 100 090 100 100 099


























































































Authors
















1083-3668201015(6)06601912$2500 Ccopy 2010 SPIE






























Inflammation 1 533


DCIS 1 194

IDC 12 22547

ILC 1 479


Adipose 9 7021

Total ROI 48 67220




































Classification




Median 875 168

Mean 130 253



[min max] [0 535] [215 953]


Accuracy 086 086 098 074 085 100 086 099 099 098

Sensitivity 090 077 087 074 009 000 077 000 000 087

Specificity 082 090 099 074 091 100 090 100 100 099


























































































Authors






























Inflammation 1 533


DCIS 1 194

IDC 12 22547

ILC 1 479


Adipose 9 7021

Total ROI 48 67220




































Classification




Median 875 168

Mean 130 253



[min max] [0 535] [215 953]


Accuracy 086 086 098 074 085 100 086 099 099 098

Sensitivity 090 077 087 074 009 000 077 000 000 087

Specificity 082 090 099 074 091 100 090 100 100 099


























































































Authors





























Classification




Median 875 168

Mean 130 253



[min max] [0 535] [215 953]


Accuracy 086 086 098 074 085 100 086 099 099 098

Sensitivity 090 077 087 074 009 000 077 000 000 087

Specificity 082 090 099 074 091 100 090 100 100 099


























































































Authors



















































































Authors


automated classification of breast pathology using local

Documents