A Bayesian Approach to Weakly-supervised Laparoscopic Image Segmentation

Introduction

This repository provides PyTorch implementation of our MICCAI2024 paper 'A Bayesian Approach to Weakly-supervised Laparoscopic Image Segmentation'.

Usage

Prepare the datasets

Download the CholecSeg8k and AutoLaparo datasets.

Extract dataset files to form folder structures like

Bayesian_WSS
├── ...
├── CholecSeg8k
│   ├── ...
│   ├── dataset
│   │   ├── CholecSeg8k
│   │   │   ├── video01
│   │   │   ├── video09
│   │   │   ├── ...
│   │   ├── target_list.txt
│   │   ├── val_samples_fold_1.txt
│   │   ├── val_samples_fold_2.txt
│   │   ├── val_samples_fold_3.txt
│   │   ├── val_samples_fold_4.txt
│   │   ├── val_samples_fold_5.txt
├── AutoLaparo
│   ├── ...
│   ├── dataset
│   │   ├── AutoLaparo_Task3
│   │   │   ├── imgs
│   │   │   ├── masks
│   │   ├── target_list.txt

Prepare the environment

Install necessary packages

pip install -r requirements.txt

Build extension module to apply DenseCRF loss

cd CholecSeg8k/utils/crfloss/wrapper/bilateralfilter
swig -python -c++ bilateralfilter.i
python setup.py install

cd AutoLaparo/utils/crfloss/wrapper/bilateralfilter
swig -python -c++ bilateralfilter.i
python setup.py install

Prepare weak annotations

For the CholecSeg8k dataset:

cd CholecSeg8k
python weak_label_simulation.py

For the AutoLaparo dataset:

cd AutoLaparo
python weak_label_simulation.py

Train and test models

With the CholecSeg8k dataset:

cd CholecSeg8k
# Default parameter values are already set.
python train.py
python inference.py

With the AutoLaparo dataset:

cd AutoLaparo
# Default parameter values are already set.
python train.py
python inference.py

Acknowledgement

We would like to express our gratitude to the following codebases:

• GBDL
• RLoss
• WSL4MIS

Note

Contact: Zhou Zheng (zzheng@mori.m.is.nagoya-u.ac.jp)