Computerized ultrasound B-scan characterization of breast nodules

The ability of computerized parameters to discriminate benign from malignant breast nodules from digitized ultrasonic acquisitions has been assessed. The images of 75 lesions, including 19 lesions proved to be malignant at histology and 56 found to be benign, were digitized and characterized by morphometric and texture parameters. The texture parameters, derived from first-order statistics, run-length matrices and co-occurrence matrices, were computed in the entire lesion and in a ring-shape region surrounding the contour of the lesion. The strongest features were found to be issued from the second region. Further investigations confirmed that the discriminant information was contained in the external part of the lesion and, to a lesser extent, in the neighboring tissue. A linear discriminant analysis using three features yielded a sensitivity of 94.7% for a specificity of 80.4% and the "leave-one-out" technique confirmed the results. Comparison with the classifications given by radiologists let us assume that information revealed by texture features is able to help the physician in reducing the number of unnecessary biopsies.