Introduction

In the frame of CEEPUS project co-operation between

two projects were started in the area of mammography in October 1999. In the first one a hierarchical decision model which evaluates data clusters of micro-calcifications (MCs) and classifies them according to severity (benign or malignant) was built. In the second project we are trying to support the physician to look at the patient as a whole with the use of breast cancer risk factors.

Analysis of mammography data (October 1999 and November 2000)

These data were got as an output from neural network for detection of MCs in mammograms, that dr. Erich Sorantin and his fellow-workers recently developed in Laboratory for digital image processing and artificial intelligence. Below one row of these data is presented:

/disk4/ausschnitte/921367ax,TIF,124 1 1 5 5 5 3,00000 2,00000 1,50000 6,00000 0,750000 1,25000 0,250000 0,750000 1,00000 0,00000 0,00000 1,40000 1,00000 0,00000 0,00000 0,00000 0,561932 4,97393 10,9739 13,9739 17,9739 725,578 25606,0 25906,0 25721,8 122,773 15073,2 662,000 864,000 737,200 86,9523 7560,70 339,000 861,000 604,400 220,000 48399,8 12 5,00000 14,0000 7,91667 2,96827 9,00000 5,00000 14,0000 7,91667 2,96827 9,00000 5,00000 14,0000 7,91667 2,96827 9,00000 0,424264 4,97393 2,08682 1,69631 2,87747 0,00000 3,00000 1,41667 0,792961 3,00000 6405 0,00187354 411,391 2,10211 0,424264 0,424264 2,42426 8 1

One row of output data from NN classification routine – typical MC cluster

Each column presents different feature of observed MC cluster. For evaluation with the hierarchical decision model the following columns were used:

41 -> number of mcs in cluster
43 -> max. (area)
44 -> mean (area)
45 -> standarddev. (area)
46 -> max - min (area)
50 -> standarddev. (perimeter)
58 -> max (distances between mcs in cluster)
60 -> standarddev. (distances between mcs in cluster)
61 -> variance (distances between mcs in cluster)
62 -> min (first shape index (i0))
67 -> area of convex hull
69 -> perimeter of convex hull
72 -> histology

For building the model the expert system shell DEX for multi-attribute decision making was used. More about it can be found at http://www-ai.ijs.si/MarkoBohanec/dex.html. The following picture presents our first model:

First hierarchical decision model for evaluation of malignancy of an MC cluster

The model is shaped as a tree of attributes. Because DEX works with discrete values of attributes (for attribute “Malignant cluster” the values are “not mal”, “maybe mal”, “prob. mal”, “likely mal” and “malignant”) a program has been developed that converts continuous numbers from source data into these values. Exclamation mark and the number in the name of an attribute tell this program which columns in source data have to be put in DEX save file as an option for evaluation. The last one attribute is known as histology for this MC cluster and is used as a quality measure of the evaluation. Evaluation of two rows (#90 and # 99) of the source data is shown below.

The first cluster is evaluated as likely malignant and it's histology marks it as malignant and the second as probably malignant with histology value of not malignant.

In the development of the model we met two main problems. The first one was how to combine attributes into a tree and how to evaluate malignancy value of this combination. With the help of experts from the field of mammography we were able to overcome this problem. But the second one was more severe and is still not resolved until present date. Its main source is DEX's ability to work only with discrete data so continuous data have to be cut into “ranges”. The main question in our case is where to put borders of these ranges. In the first model we used graphical representation of source data and figured out ranges. For solution of this problem statistics will be used. With the use of cluster analysis in the beginning of the year 2001 we will hopefully get more reliable results.

 

Breast cancer risk factors

Problems with breast cancer

Benefits from an ES

Previous work

- Ph.D. thesis from dr. Jurij Lindtner with first model in DEX
- Uroš (May 2000)
- Iztok (Nov. 2000)

Work in progress (November 2000)

- Second prototype in DEX
- Validation with known patients
- Questionnaire: List of clients, addresses needed

Future tasks (May 2001)

- Analysis of questionnaires and consequent model improvement
- Internet version of DEX and the model
- Transition of DEX and the model to PDA