Evaluating classification strategies for detection of circumscribed masses in digital mammograms

This paper reports work to investigate a computer aided diagnosis scheme for detection of circumscribed digitised masses in mammograms. The model used consists of three stages: The first segments the mammogram into regions of interest (ROIs), the second stage tries to eliminate the majority of false positives reported, and finally a third stage utilises more sophisticated processing to make a final decision on the likelihood of an ROI being a genuine mass. The paper is concerned with an evaluation of different classifiers in the difficult task of detection of masses, and the relationship between correct and incorrect detection, but also points the way to developing more sophisticated-and potentially more reliable-techniques based on the integration of multiple classifiers within a single processing structure. The study used 4 different classifiers: multivariate Gaussian classifier (MVG), radial basis function (RBF), Q-vector median (QVM), 1-nearest neighbour (1NN). Three different feature types were used: 3^rd order normalized Zernike moments, texture features, and a combined moment and textures feature vector. Various feature selection techniques have also been investigated to obtain feature sets which improve the performance of each classifier