Applying modular classifiers to mammographic mass classification

Classification is the last step in the computer aided diagnosis (CAD) system for determining whether a breast mass segmented from a digital mammogram is malignant or benign. Hence it is important to improve sensitivity at this stage. This work investigates the use of modular classifier (MoC) schemes, namely bagging and adaboost algorithms, for automated classification of mammographic masses. CAD systems containing a MoC are compared to CAD systems that contain traditional classifiers (TrC), for example single nearest mean or maximum likelihood classifiers. This study included 200 digitized mammograms, each manually segmented by a radiologist. In order to test the MoC and TrC approaches, conventional shape based features were extracted from the segmented masses. These features were then optimized using Fischer's linear discriminant analysis (LDA). When no LDA was utilized, it was observed that MoC schemes increased the sensitivity from 74% to 83% over the TrC approaches. After performing LDA, the sensitivity increased from 83% to 88% for TrC and MoC schemes, respectively.