Context This is the checkpoint for an FCN trained on the DDSM images. I am uploading it here because it would be far too slow to train without GPUs. The network was trained as a normal ConvNet and then the fully connected layers were replaced with convolutional layers so images of any size could be accepted as input. The images were sized down by half on each dimension during pre-processing, so the mean size of input images is about 2500 pixels high. Inspiration While the network performs well on the training and test data, images of size 299x299, it does not do so well when given entire scans as input, or on images from other datasets.