Breast cancer identification based on thermal analysis and a clustering and selection classification ensemble

Abstract

Breast cancer is the most common form of cancer in women. Early diagnosis is necessary for effective treatment and therefore of crucial importance. Medical thermography has been demonstrated an effective and inexpensive method for detecting breast cancer, in particular in early stages and in dense tissue. In this paper, we propose a medical decision support system based on analysing bilateral asymmetries in breast thermograms. The underlying data is imbalanced, as the number of benign cases significantly exceeds that of malignant ones, which will lead to problems for conventional pattern recognition algorithms. To address this, we propose an ensemble classifier system which is based on the idea of Clustering and Selection. The feature space, which is derived from a series of image symmetry features, is partitioned in order to decompose the problem into a set of simpler decision areas. We then delegate a locally competent classifier to each of the generated clusters. The set of predictors is composed of both standard models as well as models dedicated to imbalanced classification, so that we are able to employ a specialised classifier to clusters that show high class imbalance, while maintaining a high specificity for other clusters. We demonstrate that our method provides excellent classification performance and that it statistically outperforms several state-of-the-art ensembles dedicated to imbalanced problems. © Springer International Publishing 2013.