Fig. 2.

The proposed deep learning structure (LRP-NET) to capture changing pattern of longitudinal mammogram examinations for predicting breast cancer risk. The prediction is based on 16 images per patient. Eight CC-projections are fed into the top-left network and eight MLO projections into the bottom-left network. For each projection, 4 feature vectors belonging to the 4 corresponding priors go into the GRU model. Then the outputs of the two GRUs are concatenated to pass to the three dense layers followed by a flatten operation.