Fuzzy convex set-based pattern classification for analysis of mammographic microcalcifications

There are many different criteria for the comparative analysis of pattern classifiers. They include generalization ability, computational complexity and understanding of the feature space. In some applications such as the medical diagnostic systems it is crucial to use reliable tools, whose behavior is always predictable, so that the risk of misdiagnosis is minimized. In such applications the use of the popular feedforward backpropagation (BP) neural network algorithm can be seen as questionable. This is because it is not inherent for the backpropagation method to analyze the problem's feature space during training, which can sometimes result in inadequate decision surfaces. A novel convex-set-based neuro-fuzzy algorithm for classification of difficult-to-diagnose instances of breast cancer is described in this paper. With its structural approach to feature space the new method offers rational advantages over the backpropagation algorithm. The classification performance, computational and structural efficiencies are analyzed and compared with that of the BP network. A 20-dimensional set of "difficult-to-diagnose" mammographic microcalcifications was used to evaluate the neuro-fuzzy pattern classifier (NFPC) and the BP methods. In order to evaluate the learning ability of both methods, the relative size of training sets was varied from 40 to 90%. The comparative results obtained using receiver operating characteristic (ROC) analysis show that the ability of the convex-set-based method to infer knowledge was better than that of backpropagation in all of the tests performed, making it more suitable for use in real diagnostic systems. (C) 2001 Pattern Recognition Society. Published by Elsevier Science Ltd. All rights reserved.