Code for 'Unsupervised Surgical Instrument Segmentation via Anchor Generation and Semantic Diffusion' (MICCAI 2020).
Paper and Video Demo.
- Recommended Environment: Python 3.5, Cuda 10.0, PyTorch 1.3.1
- Install dependencies:
pip3 install -r requirements.txt.
- Download our data for EndoVis 2017 from Baidu Yun (PIN:m0o7) or Google Drive.
- Unzip the file and put into the current directory.
- The data includes following sub-directories:
image : Raw images (Left frames) from the EndoVis 2017 dataset
ground_truth : Ground truth of binary surgical instrument segmentation.
cues : Hand-designed coarse cues for surgical instruments.
anchors : Anchors generated by fusing cues.
prediction : Final probability maps output by our trained model (Single stage setting).
Simply run python3 main.py --config config-endovis17-SS-full.json .
This config file config-endovis17-SS-full.json is for the full model in the single stage setting (SS).
For other experimental settings in our paper, please accordingly modify the config file and the train_train_datadict, train_test_datadict, test_datadict in main.py if necessary.
Results will be saved in a folder named with the naming in the config file.
This output folder will include following sub-directories:
logs : A Tensorboard logging file and an numpy logging file.
models: Trained models.
pos_prob: Probability maps for instruments.
pos_mask: Segmentation masks for instruments.
neg_prob: Probability maps for non-instruments.
neg_mask: Segmentation masks for non-instruments.
Liu D. et al. (2020) Unsupervised Surgical Instrument Segmentation via Anchor Generation and Semantic Diffusion. In: Martel A.L. et al. (eds) Medical Image Computing and Computer Assisted Intervention – MICCAI 2020. MICCAI 2020. Lecture Notes in Computer Science, vol 12263. Springer, Cham. https://doi.org/10.1007/978-3-030-59716-0_63
MIT