Skip to content

Latest commit

 

History

History
96 lines (77 loc) · 4.68 KB

README.md

File metadata and controls

96 lines (77 loc) · 4.68 KB
Raw

Team Neuropoly: MICCAI 2021 MS New Lesions Segmentation Challenge

This repository contains implementations of segmentation pipelines proposed by the NeuroPoly lab for the MICCAI 2021 MS New Lesions Segmentation Challenge. The goal of the challenge was to segment the new lesions given two fluid attenuated inversion recovery (FLAIR) images (baseline and follow-up). You can also check our arXiv paper and poster.

Getting Started

This repo has been tested with Python 3.8. Follow the steps below to use this repo:

  1. Clone project: git clone https://github.com/ivadomed/ms-challenge-2021
  2. Create virtual environment and install packages: 
    cd ms-challenge-2021/
virtualenv -p python3.8 .env
source .env/bin/activate
pip install -r requirements.txt
  3. Check the sections below for dataset curation, preprocessing, and modeling. modeling/README.md provides documentation on how to run training and evaluation using the codebase.

Prerequisites

You need

installed for the preprocessing pipeline.

Check this to see how ANIMA should be installed and configured in your system.

Dataset Curation

We used the following script to curate the dataset and make it compatible with the Brain Imaging Data Structure (BIDS) specification:

python scripts/curate_msseg_challenge_2021.py -d PATH_TO_msseg_challenge_2021_DATASET

Preprocessing

The preprocessing pipeline for each subject can be found in preprocessing/preprocess_data.sh. The quality control (QC) script for this pipeline is preprocessing/qc_preprocess.py. Section 3.2 of our arXiv paper describes this pipeline in detail. You can also check our quality-control (QC) visualizations. You can find an example preprocessing visualization for a subject below.

Preprocessing steps

The preprocessing steps include:

  1. Resampling of both FLAIR sessions to isotropic 0.75mm x 0.75mm x 0.75mm resolution
  2. Spinal cord (SC) segmentation with sct_deepseg_sc on both sessions
  3. Initial registration (ses-01 -> ses-02) using sct_register_multimodal with the help of SC segmentation masks
  4. Brain extraction using bet2 on second session
  5. Finer registration (ses-01 -> ses-02) using antsRegistration
  6. Brain + SC masking on both sessions
  7. Bias-correction using animaN4BiasCorrection on both sessions
  8. Cropping of volume-of-interest (VOI)

How to run preprocessing

We are using sct_run_batch to perform preprocessing on all subjects:

sct_run_batch -path-data PATH_DATA -path-output PATH_OUTPUT -script preprocessing/preprocess_data.sh -script-args "bet2"

where PATH_DATA is the path to the BIDS data folder, and PATH_OUTPUT is where the output of preprocessing will be saved to. PATH_OUTPUT will contain data_processed and qc (among others) directories after a successful run.

Additionally, you might want to play around with -jobs and -itk-threads arguments of sct_run_batch to gain speed-ups. -jobs 16 -itk-threads 16 was what we used in joplin.

After a successful run, next step is to do quality-control (QC):

python preprocessing/qc_preprocess.py -s PATH_OUTPUT -mv

The QC script checks for:

  • whether resolutions match between the two original sessions,
  • whether all image sizes are equivalent for each subject,
  • whether isotropic-resampling (step 1.) worked as expected,
  • whether brain + SC mask leaves out any lesions from GTs (every expert and consensus),
    • NOTE: In this project, we assumed that any lesions that are not inside the brain + SC region reflects a mistake in the annotation process. Therefore, you can discard this check safely!

and outputs the following to QC i) brain + SC extraction and ii) registration:

  • per-subject visualizations
  • aggregated visualizations (all subjects)

Modeling

All modeling efforts can be found in modeling/. modeling/README.md and section 3.4 of our arXiv paper describe the deep learning architectures used in detail.

Internal use at NeuroPoly

Location of the data:

  • git+ssh://data.neuro.polymtl.ca:msseg_challenge_2021 --> this is the BIDS-converted dataset.