README.md

XAI-VesselNet-torch : Explainable AI for deep vessel segmentation

This work is a part of a PhD thesis, "Détection faiblement supervisée de pathologies vasculaires" (Weakly supervised detection of vascular pathologies).

Warnings

• This is a work in progress.
• This work was tested on Windows using a Quadro RTX 8000 GPU.

Installation

1. Clone this repository in your workspace and move into it.

2. Set PYTHONPATH environment variable :
   Windows (this will only affect the current user's environment):
   

   setx PYTHONPATH "$Env:PYTHONPATH;<workspace_path>\XAI-VesselNet-torch\xai-vesselnet".

   Linux (preferably in the .bashrc):
   

   export PYTHONPATH="<workspace_path>/XAI-VesselNet-torch/xai-vesselnet:$PYTHONPATH"

3. Create the requiered conda environment to execute XAI-VesselNet-torch

conda env create -f environment.yml

NB : Replace the <workspace_path> aliases by your own path.

Models

We integrated only a few classical models from the MONAI framework.

Model	Arg	Description
U-Net	unet	Basic U-Net. Downsampling are perform by stride, not pooling.
Residual U-Net	res-unet	Residual U-Net
Attention U-Net	attention-unet	Attention U-Net

Usage

First, activate the freshly created environment.

conda activate -n xai_vesselnet

Train a model

python ./bin/train.py [--hyperparameters HYPERPARAMETERS_JSON_PATH] MODEL TRAIN_CSV_PATH VAL_CSV_PATH

Evaluate a model

python ./bin/eval.py [--mask MASK_CSV_PATH] MODEL WEIGHTS_PATH EVAL_CSV_PATH

Infer a data

python ./bin/infer.py MODEL WEIGHTS_PATH INFER_CSV_PATH

Build a CSV dataset

We provides an easy way to define your own datasets. XAI-VesselNet-torch's scripts require CSV dataset. CSV annotations should contain pairs of image;annotation, e.g.

<absolute_path_to_image>/image_1.nii;<absolute_path_to_seg>/seg_image_1.nii
<absolute_path_to_image>/image_2.nii;<absolute_path_to_seg>/seg_image_2.nii
...
<absolute_path_to_image>/image_n.nii;<absolute_path_to_seg>/seg_image_n.nii