Development and validation of a fully automated system for detection and diagnosis of mammographic lesions

We present a comprehensive and fully automated system for computer-aided detection and diagnosis of masses in mammograms. Novel methods for detection include: selection of suspicious focal areas based on analysis of the gradient vector field, rejection of oriented components of breast tissue using multidirectional Gabor filtering, and use of differential features for rejection of false positives (FPs) via clustering of the surrounding fibroglandular tissue. The diagnosis step is based on extraction of contour-independent features for characterization of lesions as benign or malignant from automatically detected circular and annular regions. A new unified 3D free-response receiver operating characteristic framework is introduced for global analysis of two binary categorization problems in cascade. In total, 3,080 suspicious focal areas were extracted from a set of 156 full-field digital mammograms, including 26 malignant tumors, 120 benign lesions, and 18 normal mammograms. The proposed system detected and diagnosed malignant tumors with a sensitivity of 0.96, 0.92, and 0.88 at, respectively, 1.83, 0.46, and 0.45 FPs/image, with two stages of stepwise logistic regression for selection of features, a cascade of Fisher linear discriminant analysis and an artificial neural network with radial basis functions, and leave-one-patient-out cross-validation.