Fig. 4.

To learn differences between left and right breast and temporal change around a candidate location, we use a two-stream convolutional neural network (CNN). The first stream has as input a patch centered at a candidate location, the second stream a patch around the same location in either the contralateral image or the prior, using the mapping depicted in Fig. 3. All weights are shared across streams and feature maps are concatenated before the first fully connected layer.