Optimization of a breast mass classifier for computer-aided ultrasound analysis

The goal of this research was to optimize performance of a Computer-Aided Diagnostic system to identify, analyze and compare breast masses based on parameters measured in the ultrasound image. We compared case-based reasoning using Relative Similarity to an Artificial Neural Network in order to implement an objective form of the ACR BIRADS scheme to describe and score breast masses. The image feature set was reduced to nine including margins, shape, echogenicity, echo texture, orientation and posterior acoustic attenuation. Both classifiers performed well with a high ROC Az although RS performed significantly better than the ANN in Specificity, PPV and achieved the goal of very high Specificity without a reduction in Sensitivity. Compared to a preliminary version of the RS classifier this optimized version of RS has significantly higher Az (0.96 vs. 0.93).