Design and analysis of contour-independent features for classification of mammographic lesions

We present new feature descriptors specifically designed to quantify angular nonstationarity and angular dependence of pixel values in sectors of mammographic lesions. A key novelty of this work is that the proposed measures characterize the texture of masses without relying on accurate determination of their contours. An artificial neural network based on radial basis functions was used to predict the diagnosis of 120 benign masses and 26 malignant tumors in a database of full-field digital mammograms. Features were selected using stepwise logistic regression and the leave-one-patient-out method was used for cross-validation of results. An area under the receiver operating characteristic curve of 0.9890 ± 0.0114 was obtained using randomly selected centroids and an expected size of the masses. Results indicate that the use of the proposed contour-independent features can be an effective approach for computer-aided classification of mammographic lesions.