Prediction of Short-Term Breast Cancer Risk Based on Deep Convolutional Neural Networks in Mammography

This work investigated a novel framework of predicting short-term breast cancer risk by using a deep learning approach in mammography. A dataset of 675 negative screening cases were applied. 333 cases were cancer diagnosed at next screening, while 342 cases remained negative. In order to stratify these patients into high and low cancer risk group, we first used an automatically method to segment bilateral matched central regions from right and left mammography respectively. Then, three AlexNet, GoogLeNet and ResNet based deep learning models were established with ten-fold cross validation method for both difference image of bilateral matched central regions and two whole regions of bilateral breasts respectively. Using AlexNet-, GoogLeNet- and ResNet-based risk model, areas under ROC curves (AUC) are 0.56, 0.62 and 0.64 for central regions and 0.59, 0.57 and 0.65 for whole regions, respectively. When combining prediction scores of three deep learning models with a multi-agent fusion algorithm, AUCs are 0.67 and 0.67 for central regions and whole regions respectively. When fusing scores of central region-based risk model and whole region-based risk model, AUC significantly increases to 0.71 (p < 0.01). By dividing 675 cases into five subgroups based on sorting results of risk scores, the odds ratios had an significant increasing trend as the scores increased (p = 0 003). This study demonstrates feasibility of applying deep learning technology to assist investigating novel markers in mammography for helping assessment of short-term breast cancer risk and improving the efficiency of breast cancer screening in the future.