Note: the MICCAI 2018 paper only covers the multi-phase lesion detection part of this project. The liver segmentation with U-net is currently WIP.
Pixel-wise liver segmentation & multi-phase lesion detection from CT image in PyTorch.
This repo uses in-house industrial CT dataset, so the code for data loading is little nasty & unstructured, and I cannot release the dataset.
But structures of the liver segmentation data are nearly identical to 3Dircadb (http://www.ircad.fr/research/3dircadb/). And the lesion detection dataset looks like the figure 1 from our MICCAI 2018 paper.
train_liverseg_unet.py: U-net training for liver segmentation. Assumes 512x512 slices of dicom CT image and their corresponding binary segmentation masks.
ssd_liverdet/train_lesion_multiphase.py: GSSD training for liver lesion detection. Assumes 4-phase 512x512 slices of dicom CT image and their corresponding bounding box annotations.
The original SSD implementation is hard-forked from here.