A deep learning approach for efficient registration of dual view mammography

In a standard mammography study, two views are acquired per breast, the Cranio-Caudal (CC) and Mediolateral-Oblique (MLO). Due to the projective nature of 2D mammography, tissue superposition may both mask or mimic the presence of lesions. Therefore, integrating information from both views is paramount to increase diagnostic confidence for both radiologists and computer-aided detection systems. This emphasizes the importance of automatically matching regions from the two views. We here propose a deep convolutional neural network for the registration of mammography images. The network is trained to predict the affine transformation that minimizes the mean squared error between the MLO and the registered CC view. However, due to the complex nature of the breast glandular pattern, deformations due to compression and the paucity of natural anatomic landmarks, optimizing the mean squared error alone yields suboptimal results. Hence, we propose a weakly supervised approach in which existing annotated lesions are used as landmarks to further optimize the registration. To this aim, the recently proposed Generalized Intersection over Union (GIoU) is exploited as loss. Experiments on the public CBIS-DDSM dataset show that the network was able to correctly realign the images in most cases; corresponding bounding boxes were spatially matched in 68% of the cases. Further improvements can be expected by incorporating an elastic de- formation field in the registration network. Results are promising and support the feasibility of our approach.