Code for the paper entitled "3D RoI-aware U-Net for Accurate and Efficient Colorectal Cancer Segmentation"(https://arxiv.org/abs/1806.10342).
The latest codes along with weights and a test fold are now released.
Tips: a recent attempt that transfers training and inferencing to fp16 data format can further enlarge applicable volume sizes.
Here are some results of colorectal cancer segmentation, which is the case of the paper; and illustrations of another task, mandible and masseter segmentation, showing the scalability of the proposed method.
Latest experiment: simultaneously segmenting 14 organs from pelvic CTs in ~0.5s (We trained this model with 24 training samples).
