Classification of mammographic abnormalities using convolutional neural networks

Abstract

Distinguishing between benign and malignant mammography images is a complex task even for experimented radiologists and, to deal with this issue, researchers developed computer-aided diagnosis systems. This study introduces a machine learning model to classify mammography images into benign and malignant classes. We extract region of interests using the appropriate mask for each mammography image, and we feed it into a modified LeNet model. We add two parallel convolutional blocks to the original LeNet architecture, and we notice a significant increase in the performance. We pre-train the model on the DMID dataset and a subset of the BCDR dataset, and we test it on the remaining subset. The modified lightweight model reached an accuracy of 99%.