Evaluation of Learning Approaches Based on Convolutional Neural Networks for Mammogram Classification

Mammography is still considered the best screening method for detection, diagnosis and follow-up of breast cancer. A correct classification of mammographic findings demands a high expertise level of the clinician observer. For this, different Computer-aided Diagnosis systems have been developed to support the diagnosis tasks and reduce the inter or intra-observer variability caused by the complex visual information contained in mammograms. However, the classification of some findings (masses, calcifications) is still a difficult task. This work presents a methodological approach to evaluate the performance of the training process for different convolutional neural network configurations of the VGG16 Convolutional Neural Network architecture, designed to perform mammographic classification. For doing that, the impact of different learning strategies (focal loss, to deal with highly unbalance datasets, gradient clipping and learning transfer) is evaluated. The proposed method was two-fold evaluated. First, the performance for classifying between normal and abnormal Regions of Interest (ROIs) extracted from the DDSM and CBIS-DDSM datasets was explored. After that, a multi-class problem was addressed, for which a set of 5-class was included according to well-known BI-RADS classification. The obtained results reported an average accuracy of 0.92 for the binary classification and a rate of accuracy of 0.85 for the 5-class classification (with 30 epochs), reducing the convergence time (23 and 30 epochs for both binary and multi-class classification tasks, respectively).