Automated Breast Cancer Detection on Whole 3D Mammogram Using a Convolutional Neural Network Model

Breast cancer remains one of the most prevalent and deadly forms of cancer. In the United States, nearly a quarter million women are diagnosed with breast cancer annually. Twodimensional (2D) X-ray mammography provides sensitive early detection. Despite this, 2D mammography suffers from a high false-positive rate resulting in a high patient recall rate1-3. Newly emerging three-dimensional (3D) mammography produces clearer images and has been proven to improve cancer detection with lower recall rate. However, the interpretation of 3D mammogram is time consuming and requires extensive training. This demand can limit the clinical utilization of this promising technology4. Therefore, it is important to develop automated 3D mammogram interpretation tools which can aid radiologists to detect breast cancer more accurately and efficiently so that this potentially lifesaving technology can be applied more broadly and benefit more women. Objective: Convolutional neural network (CNN) is one of the most powerful deep learning (DL) tools for automatic image classification; its accuracy has surpassed almost all other traditional classification methods5. Conventional mammogram classification often requires extensive manual annotation for data training, which limits its application in real world practice. Our group has explored a variety of CNN models for classification of whole 2D mammograms to improve breast cancer detection with minimal manual annotation. To do this we built a reliable CNN classification model via transfer learning with promising preliminary results for the whole 2D mammogram. We hypothesize that a transition from 2D to 3D mammogram classification using CNN without manual intervention will improve the overall performance of automated mammogram classification. The objective of this proposal is to develop and optimize a fully automated 3D mammography classification CNN system that can analyze whole mammograms without manual image annotation.