AGSD-Surgical-Instrument-Segmentation

Code for 'Unsupervised Surgical Instrument Segmentation via Anchor Generation and Semantic Diffusion' (MICCAI 2020).

Paper and Video Demo.

Setup

• Recommended Environment: Python 3.5, Cuda 10.0, PyTorch 1.3.1
• Install dependencies: pip3 install -r requirements.txt.

Data

1. Download our data for EndoVis 2017 from Baidu Yun (PIN:m0o7) or Google Drive.
2. Unzip the file and put into the current directory.
3. 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).

Run

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.

Output

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.

Citation

TO DO.

License

MIT