Path. Res. Pract. 188,405-409 (1992)

An Expert System for Histological Typing and Grading of Invasive Breast Cancer First Set Up P. J. van Diest, J. A. M. Belien and J. P. A. Baak Institute for Pathology, Free University Hospital, Amsterdam, The Netherlands

SUMMARY

This article describes the set up of a rule-based expert system for histologic typing and grading of invasive breast cancer, which is designed to be a user-friendly tool that may be helpful for teaching and to support diagnosis making. The system raises questions and offers fixed choices to the user (usually yes/no) until a histologic diagnosis can be made with rea~onable probability or enough data are available to assign a grade. As to histologic typing, the expert system is able to make the following diagnoses: ductal carcinoma, lobular carcinoma, medullary carcinoma, colloidal carcinoma, tubular carcinoma, and invasive cribriform carcinoma. If the diagnosis "ductal carcinoma" is arrived, the system offers the option to assign a histologic grade to the lesion. A first evaluation of the system in 30 cases (five each of the different subtypes) with unequivocal diagnoses by two human experts showed that the system classified 29 of the tumours in the same way as the human experts. The discrepancy case was solved after adding one rule to the system. Ten cases where a discrepancy existed between the original diagnosis of a referring centre and a reviewing human expert were all classified by the expert system in the same way as the human expert. The expert system thus seems to perform well. Further plans for evaluating, modifying and expanding the system are disclosed.

Introduction In invasive breast cancer, several subtypes with different histologic appearance can be discriminated. The most frequently appearing subtype is the ductal type (also called "not otherwise specified" or "no special type") which accounts for approximately 80% of the invasive breast cancers 7 . About 10% is of the lobular type, and the remaining 10% is approximately equally divided over the medullary, colloidal, tubular, invasive cribriform and other types. This discrimination on the basis of histologic appearance is prognostically relevant since the medullary, colloidal (or mucinous), tubular, and invasive cribriform subtypes have a better prognosis than the lobular subtype which in turn may have a better prognosis than the ductal © 1992 by Gustav Fischer Verlag, Stuttgart

subtype 12 . However, reproducibility of histologic typing may not always be perfect 14 • Because of this, the WHO has provided uniform criteria for typing, which are described in an atlas l . Textual descriptions are, however, limited in this atlas, and no clear cut decision trees are provided. Although the very frequent ductal cancers as a group have a relatively poor prognosis, they do not form a homogeneous group. Some cases display a very good prognosis. Thus, the ductal cancers need to be further subdivided. Therefore, different grading systems 8 have been developed for ductal breast cancers. The most widely used are histologic grading4 on the basis of tubule formation, nuclear atypia and mitotic activity, and nuclear grading3 that considers only nuclear and nucleolar features. Such subjective grading systems have been proven to 0344·0338/9210188·0405$3.5010

406 . P.

J.

van Diest, J. A. M. Belien and

J.

P. A. Baak

reasonably discriminate ductal cancers with a relatively good, intermediate and bad prognosis. Reproducibility of subjective grading is, however, low5,6, 13. This may at least in part be due to the loose use of the criteria for histologic grading8, 9. The problems described above (lack of clear cut decision trees for histologic typing and loose use of criteria for subjective grading) may be overcome by using an expert system. Decision trees for histologic typing could be builtafter they have been composed by human experts - into a rule-based expert system, and criteria for histologic and nuclear grading could be integrated. Such an expert system that can nowadays be built on inexpensive personal computers 2,10 could then be of help in the diagnosis making process and also for teaching. In this article we describe the set up of an expert system for histologic typing and grading of invasive breast cancer that will serve as the basis for the long-term development of an expert system for breast cancer diagnosis. Material and Methods Criteria for Histologic Typing Since several authors have described different criteria for histologic typing, a choice has to be made as to which criteria seem relevant. Although this may at least to some extent be a matter of taste, we have found the criteria described by Page and Anderson 12 (which were to a large extent based on the criteria of the WHO! and McDivitt et a1. 11 ) to be very useful, and decided to use these for the expert system. Appendix 1 shows the microscopic criteria described by Page and Anderson for the most frequently occurring histologic subtypes of invasive breast cancer: the ductal, lobular, colloidal, medullary, tubular, and invasive cribriform subtypes.

Criteria for Histologic Grading In agreement with Bloom and Richardson4, the part of the expert system for histologic grading was based on three parameters: the amount of tubule formation, the degree of nuclear pleomorphism and the mitotic rate. For each of these parameters a score of 1 to 3 is given and the total number of points determines the histologic grade (Table 1).

The Expert System First, a decision tree was composed using a hierarchy of features, starting off with those that discriminate best between specific subtypes. The decision tree is shown in Fig. 1 (the numbers and letters in the figure correspond with the questions and diagnosis from appendix 2, respectively). This tree was then programmed using VP-Expert, a commercially available rulebased expert system shell, which was described before in detaiF. In short, using a simple word processor, an ASCII file is produced that contains a variable and value section, a rule section, a conclusion section and a question section. The variables and the values for these variables (e.g. yeslno, low/intermediatelhigh, or 0-100) have to be defined. Then, the rules that give a meaning to combinations of the values of the variables have to be created using the "IF ... THEN ... " construction, and the resulting conclusions (diagnosesf. have to be composed. Finally, the questions that ask the person consulting the expert system for a value of a variable in a user-friendly way have to be formulated. If

Table 1. Criteria for histologic grading@ (slightly modified from Page and Anderson 8 ) Parameter

Points

Tubule formation > 75% tubules with clear visible lumina 10-75% definite tubule formation, rest solid areas < 10% tubule formation

1 2 3

Nuclear pleomorphism regular nuclei, little variation in size and shape moderate variation, no extremes of size or shape marked variation, large and bizarre nuclei, multiple nucleoli Mitotic rate less than 10 mitoses per 10 HPF 10-19 mitoses per 10 HPF 2: 20 mitoses per 10 HPF

1 2 3

1 2 3

@ 3-5 points: grade I, well differentiated; 6-7 points: grade II, moderately differentiated; 8-9 points: grade III, poorly differentiated.

Table 2. Example of the construction of variables, values of variables and rules Variable

Value

Corresponding question

MUCINPOOLS

yeslno

Does the lesion contain pools of extracellular mucin?

MALCELLS

yeslno

Do the mucin pools contain islands of malignant cells?

ENTIRE

yeslno

Does the entire malignant part of the lesion consist of islands of malignant cells in mucin pools?

Rule:

IF MUCINPOOLS and MALCELLS and ENTIRE THEN diagnosis

= yes = yes = yes . =

.

mUC1110US carcmoma

the resulting file meets the (simple) syntax requirements of VP-Expert, it can directly be interpreted by the shell. Table 2 provides a real example of how this is done in practice. The system opens with the statement that the program is designed for histological typing and grading of invasive breast cancer, and then begins to ask questions. After answering each question, the system will raise a new question until a diagnosis can be made, which will then be displayed for the user with a certainty factor (CNF, 1-100) which increases as the likelihood of the diagnosis 111creases. If the diagnosis" ductal carcinoma" is arrived, the system offers the choice to either stop or on with histologic grading. If one chooses to go on with grading, the system will successively raise questions about tubule formation, nuclear pleomorphism and mitotic rate, and then assign a certain grade according to the answers given (Table 1).

Evaluation of the Typing System In order to obtain an impression of the performance of the expert system, it was evaluated in the following way: slides of 30

Breast Cancer Expert System . 407

I Y

2

~

3

n n

6

4

Y

I

13

7 Y

14

8

Y

Y

~

9

15

Y

16

17

10

mm

Y

m mm

11

mm 10

12 lIS

A B

C 0

19

E

B

21

m

F

G

10

mm Y

Y 10

m

22

Y >50

Y

B

Y

24

23

I

G

B

Y

I

26

mm 25

75% of the lesion, there is little CIS and no invasion into the surrounding tissue.

5. Lobular carcinoma Signet cells. Intracytoplasmic vacuoles. Diffuse multifocal infiltrating pattern of small, round, quite regular cells in single lines ("indian files") between collagen bundles, often growing in a targetoid periparenchymal pattern (classic type). or: Glandular cell aggregates of 20 or more cells of uniform appearance (alveolar type). or: Sheet-like pattern of irregularly shaped nests of small, noncohensive, regular cells with regular, round to oval nuclei (solid type).

6. Invasive cribriform carcinoma Stromal invasion with islands of cells resembling those seen in cribriform CIS. Individual islands have bars or arches of cells producing welldefined spaces between the cells. Often CIS with the same pattern. Regular nuclei (low to intermediate grade). Solid foci may be present if they show the same cytology.

Appendix 1: Criteria for histologic typing12

1. Ductal (or no special type) carcinoma Diagnosis by exclusion of the other types.

2. Mucinous carcinoma Pushing border between tumour and surrounding structures. Cohesive aggregates or islands with a cribriform or papillary pattern in pools of extracellular mucin. Cells lack bizarre features, nuclei show little atypia. Lesion may not contain malignant parts with another appearance.

3. Tubular carcinoma Tubular structures with round to oval, open central spaces lined by a single layer of orderly epithelium with cells appearing similar to one another. No relationship of the tubular elements to the local lobular architecture. Cells are regular and round, no nucleoli. Non-tubular areas must be less than < 10%, or the tubular areas must represent more than 2': 50% with a cribriform pattern of the non-tubular parts.

4. Medullary carcinoma Irregular bordered islands of tumour cells without sharp edges which are often interconnected. Islands do not invade th~ adjacent breast tissue but rather seem to push against it, resulting in smooth borders and no apparent capsule.

Appendix 2: Questions and Diagnosis Questions 1. "Does the lesion contain pools of extracellular mucin?"

(yes/go) 2. "Do the mucin pools contain islands of malignant cells?" (yes/go) 3. "Does the entire malignant part of the lesion consist of islands of malignant cells in mucin pools?" (yes/go) 4. "Do the mucin pools contain fibroblasts or blood vessels?" (yes/go) 5. "Does the lesion show tubule formation?" (yes/go) 6. "Does the lesion show large interconnecting islands of tumour cells?" (yes/go) 7. Is there prominent infiltration of inflammatory cells in the stroma?" (yes/go) 8. "Are these inflammatory cells present between the individual tumour cells?" (yes/go) 9. "How does the lesion relate to the surrounding tissue? " (2ushing Qorderlinvasion) 10. "What is the degree of nuclear atypia? " (!pld or !!!oderate/marked) 11. "Does more than 75% of the lesion consist of the thusfar described pattern?" (yes/go) ~2. "Are there any in situ malignant structures?" (go or ~ome/!!!any)

13. "Does the lesion contain signet cells?" (yes/go) 14. "Is a metastasis from another site excluded?" (yes/go) 15. "Do the tumour cells contain intracytoplasmic vacuoles? " (yes/go)

Breast Cancer Expert System . 409 16. "Do the tumour cells form indian files?" (yes/go) 17. "Do these indian files grow in a targetoid pattern?" (yes/go) 18. "Do the tumour cells grow in an alveolar pattern?" (yes/go) 19. "Do the tumour cells grow in sheets?" (yes/go) 20. "Do these tubules consist of round to oval, single layers of similar appearing cells?" (yes/go) 21. "Do the cells form so-called cribriform structures, islands having bars or arches and containing lumina?" (yes/go) 22. "Does the lesion also contain solid malignant foci?" (yes/go) 23. "Do these solid foci show the same cytologic pattern as the cribriform structures?" (yes/go) 24. "Do these tubular structures lie individually dispersed in abundant stroma?" (yes/go) 25. "Does the lesion also contain non-tubular areas?" (yes/go) 26. "Assess the percentage non-tubular areas" (0 :S % :S 100)

27. "Do the non-tubular areas have a cribriform pattern?" (yes/go)

Diagnoses A B C D E F

= mucinous carcinoma

= ductal carcinomas = mammary lesion with benign mucin pools

= mammary lesion with edematous stroma

=

medullary carcinoma

= exclude metastasis from another site, then restart consul-

tation G = lobular carcinoma H = invasive cribriform carcinoma I = tubular carcinoma

References 1 Azzopardi JG, Chepieck OF, Hartmann WH, Jafarey NA, L1ombart-Bosch A, Ozzello L, Rilke F, Sasano N, Sobin LH, Sommers SC, Stalsberg H, Sugar J, Williams AO (1981) Histologic Typing of Breast Tumors, 2nd edition. World Health Organization, Geneva

2 Baak JPA, Kurver PHJ (1988) Development and use of a rule-based pathology expert consultation system. Analyt Quant Cytol HistollO: 214-218 3 Black MM, Opler SR, Speer FD (1955) Survival in breast cancer cases in relation to structure of the primary tumor and lymph nodes. Surg Gynecol Obstet 100: 543-551 4 Bloom H]G, Richardson WW (1957) Histological grading and prognosis in breast cancer. Br] Cancer 11: 359-377 5 Cutler SJ, Black MM, Friedell GH, Vidone RA, Goldenberg IS (1966) Prognostic factors in cancer of the female breast II. Reproducibility of histopathologic classification. Cancer 19: 75-82 6 Delides GS, Garas G, Georgouli G,]iortziotis D, Lecca], Liva T, Elemenoglou] (1982) In trala bora tory variations in the grading of breast carcinoma. Arch Pathol Lab Med 106: 126-128 7 Diest P] van, Baak ]PA, Galen C van, Matze-Cok P, Wisse-Brekelmans ECM, Beek MWPM van, Bellot SM, Fijnheer ], Gorp LHM van, Los], Peterse ]L, Ruitenberg HM, Schapers RFM, Schipper MEI,Somsen]G, Willig APW, AriensATh (1990) Reproducibility of mitosis counting in 2469 breast cancer specimens: Results from the Multicenter Morphometric Mammary Carcinoma Project. Hum Pathol, in press 8 Elston CW (1987) Grading of invasive carcinoma of the breast. In: Diagnostic Histopathology of the Breast. Page DL, Anderson T] (Eds) Churchill Livingstone, Edinburgh 9 Elston CW, Gresham GA, Rao GS, Zebro T, Haybittle ]L, Houghton] (1982) The cancer research campaign (Kings/Cambridge) trial for early breast cancer - pathological aspects. Br] Cancer 45: 655-669 10 Ginneken AM van (1989) Interactive consultation and formalization of knowledge applied to ovarian pathology. Thesis, Erasmus University, Rotterdam, The Netherlands 11 McDivitt RW, Stewart FW, Berg]W. Tumors of the breast. Atlas of tumor pathology, second series, fascicle 2. Armed Forces of Pathology, Washington DC, USA 12 Page DL, Anderson T] (1987) Diagnostic Histopathology of the Breast. Churchill Livingstone, Edinburgh 13 Stenkvist B, Westman-Naeser S, Vegelius ], Holmquist ], Nordin B, Bengtsson E, Eriksson 0 (1979) Analysis of reproducibility of subjective grading systems for breast carcinoma. ] Clin Pathol 32: 979-985 14 Uyterlinde AM, Baak ]PA, Schipper NW, Peterse ]L, Meijer ]WF, Vooys GP, Matze-Cok P (1990) Malignant behaviour of invasive breast cancers detected with population screening. Comparison of prognostic value of morphometric and flow cytometric features with a hospital group of patients. Int] Cancer 48: 173-181, 1991

Received September 27, 1991 . Accepted October 28, 1991

Key words: Breast cancer - Typing-grading - Expert system Prof. Dr. ]. P. A. Baak, Department of Pathology, Free University Hospital, P.O. Box 7057, NL-I007 MB Amsterdam, The Netherlands