A U-Net is an end-to-end segmentation network that should work reasonably well in this low data setting. It will take an image, progressively convolve it to a collection of many small feature maps, and then progressively up-convolve the maps while combining with crops from the previous layers.
