This repository contains the code of the paper Interpretability-guided Data Augmentation for Robust Segmentation in Multi-centre Colonoscopy Data, accepted at the MLMI workshop at MICCAI 2023.

A pre-print of the paper is available at this link.

All experiments were carried out using the following:

albumentations==1.3.0
captum==0.6.0
hydra-core==1.3.2
kornia==0.6.12
lightning==2.0.0
monai==1.1.0
numpy==1.23.0
omegaconf==2.3.0
pandas==1.4.3
psutil==5.9.4
scikit-image==0.19.2
scikit-learn==1.1.1
torch==1.13.1
torchmetrics==0.11.4
torchvision==0.14.1
wandb==0.13.9

All experiments were conducted on the open-source PolypGen dataset. You can ask for access to the dataset at this link.

To launch the training of one of the available models run the following:

python /PATH_TO_REPO/train.py 

All configuration parameters are handles by Hydra, to run different experiments modify the config files available in the folder conf. Please refer to Hydra docs to see how to over-ride configuration parameters directly from the command line.

• Remove artifact creation in W&B logging

• Thanks to spthermo for the original Pytorch implementation of SDNet.
• Thanks to sharib-vision for the original Pytorch implementations of UNet and DeepLabV3+.
• Thanks to DebeshJha for the original dataset repository.