COMPUTERIZED SCENE SEGMENTATION FOR THE DISCRIMINATION OF ARCHITECTURAL FEATURES IN DUCTAL PROLIFERATIVE LESIONS OF THE BREAST

, . 181: 374–380 (1997)

COMPUTERIZED SCENE SEGMENTATION FOR THEDISCRIMINATION OF ARCHITECTURAL FEATURESIN DUCTAL PROLIFERATIVE LESIONS OF THE

BREAST

. 1, . 1*, . 2, 2, 3 . 1

1Department of Pathology, The Queen’s University of Belfast, Belfast, Northern Ireland, U.K.2Optical Sciences Center, University of Arizona, Tucson, Arizona, U.S.A.

3Institute of Pathological Anatomy and Histopathology, University of Ancona, Ancona, Italy

SUMMARY

The distinction between ductal hyperplasia (DH) and ductal carcinoma in situ (DCIS) still remains a problem in the histologicaldiagnosis of non-invasive breast lesions. In this study, a method was developed for the automatic segmentation and quantitative analysisof breast ducts using knowledge-guided machine vision. This permitted duct profiles and intraduct lumina to be identified and their shape,size, and number computed. These were used to derive measures of duct cribriformity and architectural complexity which were examinedas an objective tool in the characterization of duct pattern in proliferative lesions. A total of 215 images of ducts were digitally capturedfrom 22 cases of DCIS and 21 cases of DH diagnosed independently by two pathologists. The cribriformity index proved to be a usefulmeasure of duct architecture, showing a monotonic increase with increasing duct complexity. The number of lumina also increased withincreasing overgrowth of ductal epithelium until the duct was filled. Discriminant analysis of the duct characteristics for benign andmalignant groups selected the lumen area/duct area ratio and the duct area as significant discriminatory variables and they werecombined into a discriminant function. Of the lumen features, the mean area of the lumen and the polar average (mean of the distributionof the number of events with an increasing spiral from the centre of the duct) were combined into a second discriminant function. Plottingcases against these two functions provided good separation of DH and DCIS groups, with correct classification estimated on the trainingsample as being over 80 per cent. With an increasing incidence of complex proliferative lesions arising from mammography, the abilityto diagnose these lesions correctly is more important than ever. The use of expert system-guided machine vision facilitates thequantitative evaluation of breast duct architecture; along with established histological and cytological criteria, it is hoped that this willlead to a more objective means of diagnosis and disease classification. ? 1997 by John Wiley & Sons, Ltd.

J. Pathol. 181: 374–380, 1997.No. of Figures: 5. No. of Tables: 7. No. of References: 15.

KEY WORDS—hyperplasia; ductal carcinoma in situ; breast adenocarcinoma; machine vision; histometry; quantitative pathology

INTRODUCTION

The diagnosis of ductal hyperplasia (DH) and ductalcarcinoma in situ (DCIS) still remains a problem in thehistological diagnosis of breast lesions. With the increas-ing frequency of detection of such abnormalities due tomammography, the ability to diagnose these lesionscorrectly and determine the appropriate managementfor individual patients is more important than ever.However, the established criteria for distinguishing DHfrom DCIS are subjective and include features suchas the architectural pattern of ducts, the presence ofnecrosis, the pattern of nuclear spacing and orientation,

and the presence of myoepithelial cells. Of considerableinterest is the analysis of duct architecture which showsa close association with the nature of the epithelialproliferation.1 The identification of these patterns con-tributes to the identification of the histological catego-ries of mild, moderate, and florid DH and DCIS. It isclear that these groups form a spectrum of histologicalchange; a proportion of cases fall into the intermediatecategory of atypical ductal hyperplasia (ADH) andwhile specific descriptive rules have been defined,1 thereproducibility of these diagnoses is variable, with somestudies showing good agreement2 and others showingpoor reproducibility.3Recently, the automatic segmentation and morpho-

metric analysis of complex histological images have beenmade possible using knowledge-guided machine vision.4This model-based approach uses a knowledge file5 whichcomprises a comprehensive list of all the diagnosticterms, interpretative transforms, functions, definitions,constraints, and procedures necessary to segment andevaluate a given histological scene. The knowledge file isspecifically designed for a particular type of histological

*Correspondence to: Dr Peter W. Hamilton, Quantitative Pathol-ogy Laboratory, Department of Pathology, The Queen’s University ofBelfast, Grosvenor Road, Belfast BT12 6BL, Northern Ireland, U.K.

Contract grant sponsor: Ulster Cancer Foundation.

Contract grant sponsor: The Pathological Society.

Contract grant sponsor: NIDEVR.

Contract grant sponsor: National Institute of Health, Bethesda,Maryland; Contract grant number: R35CA 538701.

CCC 0022–3417/97/040374–07 $17.50 Received 12 April 1996? 1997 by John Wiley & Sons, Ltd. Accepted 7 October 1996

scene. As image segmentation is underway, segmentedobjects are tested against constraints defined in theknowledge file. If they do not meet specific criteria, theyare subjected to further segmentation until they can beuniquely assigned to a given histological component.This is achieved using ‘locally adaptive segmentation’,where a variety of segmentation algorithms may beapplied in different locations of the same image toresolve segmentation problems. Once the segmentedcomponents of the histological scene are identified,they are recombined to form identifiable architecturalstructures such as glands.This knowledge-guided approach has been applied

with success to the analysis of glandular tissues in thecolon6 and prostate,7,8 with the goal of automaticallyidentifying and measuring the glandular componentswithin the image. In the colon, this permitted themeasurement of a nuclear stratification index for thegrading of dysplasia9 and in the prostate, the compu-tation of a cribriformity index.8The objective of this study was to adapt the existing

machine vision system to enable the segmentation andanalysis of breast ducts showing DH and DCIS. Acribriformity index was defined and this, together withother measures relating to the ductal architecture, wasexamined as an objective tool in the characterization ofduct pattern and as a possible means of distinguishingDH from DCIS.

MATERIALS AND METHODSPatientsThe biopsies studied were all from women who had

mammographic abnormalities (either mass lesions,microcalcification, or both) detected as part of theNational Health Service (U.K.) Breast Screening Pro-gramme and all of the women were aged 50–65 years.Localization biopsies were carried out and the specimenswere X-rayed to ensure that the lesion had beenremoved. The specimen was then fixed overnight inunbuffered formal saline. After fixation, tissue blockswere taken from the area of the mammographic abnor-mality and these were routinely processed before embed-ding in paraffin wax. Four-micrometre sections were cutand stained with haematoxylin and eosin.The slides from each case were independently

reviewed by two experienced pathologists (NA, JMS)and were categorized as DCIS or DH using the diagnos-tic criteria outlined by Page and Anderson.1 Examples ofcomedo DCIS were excluded from the study. Cases wereonly included in the study where both pathologistsarrived at the same diagnosis. The DCIS group thuscomprised 22 cases and the DH group 21 cases.

Image capture

For each case, images of five ducts were collected at20:1 objective magnification. Images were recorded on avideophotometer equipped with a 3-chip CCD camera(SONY) and a 0·5 ìm precision co-ordinate stage. Anarrow band contrast filter with transmission centredaround 620 nm was inserted.

Knowledge-guided image analysis

The software developed for the segmentation of breastduct structures was based on previous programs for theanalysis of colonic and prostatic tissues.6–8 Identificationof ducts was based on scene segmentation using aknowledge-guided approach. Several interacting expertsystems controlled the process.

Quantitative features and cribriformity index

Once the duct had been identified, a number ofquantitative measurements were made from the ductcomponents. For each duct, five features were com-puted. These are listed in Table I. A cribriformity index(feature 1) was defined as before7 as the form factor(shape index) of the lumen divided by the form factor ofthe gland outline. Feature 4 was based on the distanceratio which is illustrated in Fig. 1.In addition, each lumen space was characterized by

seven features listed in Table II. The polar averagemeasurements were calculated from distributions ofthe number of events (lumen) encountered with anincreasing spiral from the centre of the duct.

Table I—List of the glandular features computed

Feature 1. Lumen form factor/gland outline form factor(cribriformity index)

Feature 2. Total lumen area/gland area.Feature 3. Gland area/10 000Feature 4. Average distance ratio (see Fig. 1)Feature 5. Number of lumen

Fig. 1—Illustration of the distance average calculated for each lumen.d1 represents the distance from the duct centre to the lumen centre. d2represents the distance from the lumen centre to the nearest point onthe duct outline chaincode

Table II—List of the lumen features computed

Feature 6. Form factorFeature 7. Nearest distance ratioFeature 8. AreaFeature 9. Number of polar average bins having dataFeature 10. Centre bin of polarized average dataFeature 11. Mode of polarized average dataFeature 12. Largest distance ratio

375IMAGE ANALYSIS OF BREAST LESIONS

? 1997 by John Wiley & Sons, Ltd. J. Pathol. 181: 374–380 (1997)

Fig. 2—A successfully segmented breast duct with delineated lumen boundaries and duct profile

Fig. 3—The changes in the numerical values of the cribriformity index and lumen number with increasing duct complexity

376 N. H. ANDERSON ET AL.


Statistical analysis

From the series, images were selected to representsome of the distinct patterns seen in ductal lesions. Thequantitative values obtained for these were compared.Stepwise discriminant analysis was used to select

features which best discriminated the morphologicallypredefined DCIS and DH groups. This was carried outinitially by (1) treating the duct images as independentobservations and (2) averaging the image data withinpatients to assess discrimination between cases. In thediscriminant analysis, the entry level was set to F=3·0 toreduce the chance of spurious feature selection. Selectedfeatures were combined into a linear discriminant func-tion and the percentage of correctly classified cases wascalculated. In addition, a non-linear Bayesian decisionboundary was calculated as a means of improving thecorrect classification rate.

RESULTS

Scene segmentationThe segmentation of breast duct images was success-

ful in 86 per cent of cases (Fig. 2). Segmentation tookplace in five distinct phases.1. The initial segmentation of the gland was carried

out by adjusting interactively the grey level threshold.This permitted the identification of the duct structure,distinguishing it from the background stroma. Thebackground was then allocated a low grey value to avoidconfusion in the identification of lumen in the next step.2. The next processing step segmented all objects with

a grey value less than that set for the background. Thisidentified ‘empty’ areas within the duct profile whichmight be luminal spaces. The knowledge file defined thecriteria for the definition of such objects and eachsegmented object was tested against these criteria. If theconditions were satisfied, the segmented objects wereaccepted as gland lumina. The lumen outlines were thenFourier-transformed to provide a smoothed profile.3. Next; a grey level threshold was set to identify the

basement membrane border of the duct. This wasFourier transformed to create a smoothed boundary.4. In order to reconstruct the components of the

scene, the topology of the scene was then scrutinized bythe expert system. Luminal spaces were identified asbelonging to a single duct profile based on whether ornot they were enclosed within ductal epithelial segments.In cribriform glands, several lumen spaces might belongto the same gland and this was defined in the knowledgefile.5. Once complete ducts with their associated lumina

had been identified and reconstructed (Fig. 2), theycould then undergo processing for measurement ofquantitative features.

Measures of gland complexity

One of the main objectives of this study was to devisefeatures which provide a useful measure of duct com-plexity. The cribriformity index proved to be useful as it

increased with increasing duct complexity (Fig. 3). Thenumber of lumina also increased with increasing over-growth of the ductal epithelium until the duct was filled.The other features were less useful as they did not showa monotonic increase with increasing hyperplasia.

Statistical analysis

The means and standard deviations for the measuredfeatures are shown in Table III.Discriminant analysis of the duct characteristics for

DH and DCIS groups selected the lumen/duct area ratio(feature 2) and the duct area (feature 3) as significantdiscriminatory variables. The same features wereselected for both the image-based and the patient-baseddiscrimination. When combined into a discriminantfunction, they provided a correct image classification of63 per cent (Table IV) and a correct patient classificationof 65 per cent (Table V). As the defined linear thresholdbetween benign DH and malignant DCIS cases is adiscriminant score of zero, the rule is only useful if casesshow positive discriminant function scores, as few

Table III—Means (standard deviations) of the feature valuesfor the two diagnostic groups

Feature DH (n=21) DCIS (n=22)

1 0·27 (0·18) 0·28 (0·43)2 0·15 (0·08) 0·22 (0·11)3 9·83 (5·00) 15·24 (9·93)4 2·29 (1·27) 2·70 (0·94)5 6·09 (3·66) 8·64 (4·26)6 0·63 (0·09) 0·63 (0·08)7 0·35 (0·11) 0·32 (0·05)8 1734·80 (879·75) 2365·99 (662·09)9 3·47 (1·70) 3·44 (1·38)10 10·21 (0·51) 10·36 (0·53)11 4·94 (1·33) 5·40 (1·52)12 2·46 (1·65) 2·70 (0·95)

Table IV—Image classification table for discriminant functionbased on duct features with 63 per cent of cases correctlyclassified

Benign Malignant

Benign 65 (75%) 22 (25%)Malignant 44 (47%) 50 (53%)

Table V—Patient classification table for discriminant functionbased on duct features with 65 per cent of cases correctlyclassified

Benign Malignant




benign cases show positive scores whereas almost halfof the images/patients diagnosed as malignant havepositive scores (Table V).Morphometric data for individual lumina were aver-

aged for each image and subjected to discriminantanalysis between DH and DCIS groups. Here the meanarea of the lumen (feature 8) and the number of polar

average bins with data (feature 9) were selected asdiscriminatory variables. Again, the same features wereselected for both the image-based and the patient-baseddiscrimination. In combination, a linear classificationrule was able to achieve a correct image classificationrate of 70 per cent (Table VI) and a correct patientclassification of 72 per cent (Table VII).A scatterplot of the cases against the two discriminant

functions for images is shown in Fig. 4. This illustratesseparation of the two diagnostic groups in feature space,although extensive overlap is still evident. A scatterplotof the patient scores is shown in Fig. 5, which shows amuch better separation of DCIS and DH. Calculation ofa non-linear Bayesian decision boundary based on theprobability density functions of the two sets of discrimi-nant score data allowed an even better correct patientclassification rate of 81 per cent.

DISCUSSION

Until recently, automated analysis of histologicalimagery was considered a major problem. While it stillremains a complex task, it is made possible through theuse of a knowledge-guided expert system. The develop-ment of such a system relies on transforming one’s visualinterpretative skills into a logical set of rules, functions,

Table VI—Image classification table for discriminant functionbased on lumen features with 70 per cent of cases correctlyclassified

Benign Malignant


Table VII—Patient classification table for discriminant func-tion based on lumen features with 72 per cent of cases correctlyclassified

Benign Malignant


Fig. 4—Scatterplot of DH (closed circles) and DCIS (open circles) images against the two discriminant functions for duct and lumenfeatures. Some separation of groups but with extensive overlap can be seen. The 95 per cent confidence ellipses are drawn, showing asignificant different between the two diagnostic groups



routines, and processing sequences which enable imageunderstanding by the machine. By image understanding,we mean that the machine has the ability to identify(segment) the important histological componentswithin an image and knows which components containdiagnostically useful information. Conventional descrip-tive diagnostic clues must then be converted intomeasurements which can be computed from the imagecomponents.The architectural appearance of breast ducts has been

advanced as a useful diagnostic feature in the classifi-cation of intraductal proliferations in the breast. In thisstudy, a cribriformity index and other measures relatingto the appearance of breast ducts and their lumina weredevised as metrics to assess the changes seen in breastduct architecture. The cribriformity index was shownadequately to map increased duct complexity, fromsingle lumen to multilumen ducts, but was poor atdiscriminating DH from DCIS. This was not un-expected. Varying architectural types can be associatedwith both benign and malignant ducts and even visuallythe basic gland/lumen shape is not sufficient to distin-guish clearly between benign and malignant lesions.Other features relating to the area of ducts and lumen

and the position of lumina within the duct as describedby polar averaging were found to be better discrimina-tory features. The use of lumen characteristics appeared

to provide the best discrimination of benign and malig-nant groups, although the correct classification ratewhen duct and lumen data were examined separately didnot exceed 72 per cent. However, by combining duct andlumen data in a bivariate scatterplot of discriminantscores, a better separation of diagnostic groups wasachieved. The calculation of a Bayesian decision bound-ary for the patient data significantly increased the per-centage of correctly classified cases to 81 per cent.Obviously, such classification rules need to be validatedon a larger set of cases not used in the derivation of therule. Nevertheless, these results illustrate that quanti-tative data obtained from automated segmentation ofbreast ducts could be used to improve the objectivityof diagnostic classifications in breast pathology.Other types of classification strategy can be sought.

The histopathological diagnosis of DH as opposed toDCIS in this study was not based solely on duct archi-tecture, but also on the other common features used inthe differential diagnosis of these conditions. For thisreason, the duct images associated with an individualpatient did not always show a homogeneous architec-tural pattern. The pooling of data within patients couldtherefore result in a poor correct classification into be-nign and malignant groups. An alternative approach wasinvestigated where the individual duct data were retainedand patients were classified on the basis of whether they

Fig. 5—Scatterplot of DH (closed circles) and DCIS (open circles) patients against the two discriminant functions. Smoothing of thedata by averaging improves the separation of individual patients in feature space. The 95 per cent confidence ellipses illustrate thatmorphometric data for the two groups are significantly different. Drawing a Bayesian decision boundary provides a correct classificationof 81 per cent of cases



demonstrated more than two ducts showing malignantcharacteristics in numerical terms (discriminant scoregreater than zero). Here, a correct classification of up to81 per cent, based on duct measurements alone, could beobtained. This concept matches the recommendedcriteria for conventional histological diagnosis: that atleast two glands should show the features of DCISbefore such a diagnosis is made.1It is clear that ductal characteristics carry useful diag-

nostic information for the discrimination of patients withDH and DCIS. However, duct architecture alone is notsufficient to identify and discriminate DH and DCIS andthe use of other features is necessary. Nuclear featuressuch as orientation, pleomorphism, and distribution andthe presence of swirls, necrosis, and Roman Bridges areconventionally used as additional clues in the differentialdiagnosis of DH and DCIS. Efforts are underway toderive additional quantitative data from these segmentedglands which provide data on nuclear orientation andspatial distribution which can be combined with ductarchitectural features. This should provide betterdiscrimination between hyperplasia and DCIS.Many of the features used in defining intraductal

proliferations in the breast do not lend themselves easilyto quantitative evaluation. In such circumstances, it maybe better to use a descriptive classifier. Bayesian beliefnetworks represent a useful way to represent descriptiveknowledge and to handle uncertainty in the assessmentof visual clues. In pathological decision-making, theyhave already been applied in breast cytology10,11 and inthe classification of prostate neoplasia.12–15 This group iscurrently developing such an approach for the classifi-cation of intraductal breast lesions. It is hoped that this,in combination with quantitative data, will improvethe consistency, reproducibility, and accuracy of DCISdiagnosis in breast histopathology.

ACKNOWLEDGEMENTS

This work was supported by grants from the UlsterCancer Foundation, the Pathological Society, and

NIDEVR (PWH/NHA), and by a grant (R35CA 538701to PHB) from the National Institute of Health,Bethesda, Maryland to the National Cancer Institute.The content of this paper is solely the responsibility ofthe authors and does not necessarily represent theofficial views of the National Cancer Institute.

REFERENCES

1. Page DL, Anderson TJ. Diagnostic Histopathology of the Breast.Edinburgh: Churchill Livingstone, 1987; 120–192.

2. Schnitt SJ, Connolly JL, Tavasoli FA, et al. Interobserver reproducibility inthe diagnosis of ductal proliferative lesions of the breast. Am J Surg Pathol1992; 16: 1133–1143.

3. Rosai J. Borderline epithelial lesions of the breast. Am J Surg Pathol 1991;15: 209–221.

4. Bartels PH, Bartels HG, Shoemaker RL, Thompson D. Machine visionssystem for diagnostic histopathology. Pathol Res Pract 1989; 185: 635–646.

5. Bartels PH, Thompson D, Weber JE. Construction of a knowledge file foran image understanding system. Pathol Res Pract 1992; 188: 396–404.

6. Thompson D, Bartels PH, Bartels H, Hamilton PW, Sloan JM. Knowledge-based segmentation of colorectal histologic imagery. Anal Quant CytolHistol 1993; 15: 236–246.

7. Bartels PH, Thompson D, Bartels HG, Montironi R, Scarpelli M, HamiltonPW. Machine vision-based histometry of premalignant and malignantprostatic lesions. Pathol Res Pract 1995; 191: 935–944.

8. Thompson D, Bartels PH, Bartels HG, Montironi R. Image segmentationof cribriform gland tissue. Anal Quant Cytol Histol 1995; 17: 314–322.

9. Hamilton PW, Bartels PH, Sloan JM, Thompson D. Knowledge guidedsegmentation and morphometric analysis of colorectal dysplasia. AnalQuant Cytol Histol 1995; 17: 172–182.

10. Hamilton PW, Anderson N, Bartels PH, Thompson D. Expert systemsupport using Bayesian belief networks in the diagnosis of fine needleaspiration biopsy specimens of the breast. J Clin Pathol 1994; 47: 329–336.

11. Hamilton PW, Anderson N, Bartels PH, et al. An interactive decisionsupport system for breast cytology. Anal Quant Cytol Histol (in press).

12. Bibbo M, Minimo C, Xiao J, et al. A workstation for objective grading oftumours. In: Wied GL, Bartels PH, Rosenthal DL, Schenk U, eds.Compendium on the Computerised Cytology and Histology Laboratory.Chicago: Tutorials of Cytology, 1994; 89–95.

13. Bibbo M, Bartels PH, Pfeifer T, Thompson D, Minimo C, Galera DavidsonH. Belief network for grading prostate lesions. Anal Quant Cytol Histol1993; 15: 124–135.

14. Montironi R, Bartels PH, Thompson D, Scarpelli M, Hamilton PW.Prostatic intraepithelial neoplasia. Development of a Bayesian beliefnetwork for diagnosis and grading. Anal Quant Cytol Histol 1994; 16:101–112.

15. Montironi R, Bartels PH, Thompson D, Scarpelli M, Hamilton PW.Prostatic intraepithelial neoplasia (PIN). Performance of Bayesian beliefnetwork for diagnosis and grading. J Pathol 1995; 177: 153–162.



Documents

COMPUTERIZED SCENE SEGMENTATION FOR THE DISCRIMINATION OF ARCHITECTURAL FEATURES IN DUCTAL PROLIFERATIVE LESIONS OF THE BREAST