Adaptive Splitting and Selection ensemble for breast cancer malignancy grading

Abstract

The article presents an application of Adaptive Splitting and Selection (AdaSS) ensemble classifier in a real-life task of designing an efficient clinical decision support system for breast cancer malignancy grading. We approach the problem of cancer detection form a different angle - we already know that a given patient has a malignant type of cancer and we want to asses the level of that malignancy to propose the most efficient treatment. We carry a cytological image segmentation process with fuzzy c-means procedure and extract a set of highly discriminative features. However, the difficulty lies in the fact, that we have a high disproportion in the number of patients between the groups, which leads to an imbalanced classification problem. To address this, we propose to use a dedicated ensemble model, which is able to exploit local areas of competence in the decision space. AdaSS is a hybrid combined classifier, based on an evolutionary splitting of object space into clusters and simultaneous selection of most competent classifiers for each of them. To increase the overall accuracy of the classification, in the hybrid training algorithm of AdaSS we embedded a feature selection and trained weighted fusion of individual classifiers based on their support functions. Experimental investigation proves that the introduced method is more accurate than previously used classification approaches.