Automatic classification of breast tissue according BIRADS categories using a clustering approach

1. Purpose
Recent studies showed that the sensitivity of CAD systems is significantly decreased as the density of the breast increases [1]. This paper proposes a new automatic methodology to classify breasts according their tissue density into BIRADS categories.

2. Methods
Once the breast profile is segmented [2], we use the Fuzzy C-Means algorithm to group those pixels from similar tissue into two separate categories: normal and dense tissue. To this proposal, we initialized the respective seeds with the grey level values that represent 15% and 85% of the accumulative histogram. When FCM finishes, a set of morphological and texture (co-occurrence matrices based) features for both classes are extracted. Finally, using the k-Nearest Neighbours and the ID3 Decision Tree classifiers, the breast is classified according BIRADS categories.

3. Results and conclusions
The method was tested over a set of 320 mammograms taken from the MIAS database using a leave-one-woman-out method, in which each sample is analysed by a classifier which is trained using all other samples except for those from the same woman. Results obtained shown that both classifiers almost reach 50% of well-classified mammograms into the four BIRADS categories. However, when reducing the dimensionality of this four-classification problem into the two-class one: low density/high density, most of the bad-classified mammograms are well-classified, reaching more than 80% of correct classification. Moreover, results shown that high density breasts are most correctly classified using the k-Nearest Neighbours, while ID3 Decision Tree performs better for low dense breasts. Further work is focused on how to combine both classifiers in order to take advantage of this fact.