Classifiers trained on dissimilarity representation of medical pattern: A comparative study

In this paper we investigate the feasibility of some typical techniques of pattern recognition for the classification of medical examples. The learning of the classifiers is not made in the traditional features space but it can be made by constructing decision rules on dissimilarity (distance) representations. In such a recognition process a new object is described by its distances to (a subset of) the training samples. Purpose of this work is the development of an automatic classification system which could be useful for radiologists in the investigation of breast cancer. The software has been designed in the framework of the MAGIC-5 collaboration. In the automatic classification system the suspicious regions with high probability to include a lesion are extracted from the image as regions of interest (ROIs). Each ROI is characterized by some features extracted from co-occurrence matrix containing spatial statistics information on ROI pixel gray tones. A dissimilarity representation of these features is made before the classification. A Feed-Forward Neural Network (FF-NN), a K-Nearest Neighbour (K-NN) and a Linear Discriminant Analysis (LDA) are employed to distinguish pathological records from not-pathological ones by the new features. The results obtained in terms of sensitivity (percentage of pathological ROIs correctly classified) and specificity (percentage of healthy ROIs correctly classified) will be comparatively presented. The K-NN classifier gives slightly better results than FF-NN and LDA accuracy (percentage of cases correctly classified) on two-classes problem (pathologic or healthy patients).