Catalyst.Neuro [](https://join.slack.com/t/catalyst-team-core/shared_invite/zt-d9miirnn-z86oKDzFMKlMG4fgFdZafw) [](https://github.com/catalyst-team/neuro/graphs/contributors)

This repository provides a 3D Brain Segmentation Pipeline for a structural magnetic resonance imaging (MRI) scan using the Catalyst framework using the Mindboggle Dataset as an example. It also provides pre-trained models for Gray Matter White Matter (GMWM) segmentation, 104 class brain atlas segmentation, and 31 class brain atlas segmentation and with usage shown in example notebooks. Segmenting a structural MRI is an important processing step that enables subsequent inferences about tissue changes in development, aging, and disease. This work is based on the following papers: An (almost) instant brain atlas segmentation for large-scale studies and End-to-end learning of brain tissue segmentation from imperfect labeling.

Developed in a partnership with

Pretrained Model Statistics

Model	Classes	# of Brains (Training)	# of Brains (Validation)	# of Brains Test	Macro DICE
MeshNet GMWM	3	20	4	4	.9565
MeshNet Dropout GMWM	3	20	4	4	.8748
MeshNet Large GMWM	3	20	4	4	.9652
MeshNet Large GMWM	3	20	4	4	.9624
UNet GMWM	3	20	4	4	.9624
MeshNet Large Mindboggle	31	70	10	20	.6742
UNet Mindboggle	31	70	10	20	.6771
MeshNet Large HCP Atlas	104	770	27	100	~.85

Model	Inference Speed	Model Size
MeshNet GMWM	116 subvolumes/sec	.89 mb
MeshNet Dropout GMWM	115 subvolumes/sec	.89 mb
MeshNet Large GMWM	19 subvolumes/sec	9mb
MeshNet Large Dropout GMWM	19 subvolumes/sec	9mb
UNet GMWM	13 subvolumes/sec	288 mb
MeshNet Large Mindboggle	19 subvolumes/sec	9 mb
UNet Mindboggle	13 subvolumes/sec	288 mb
MeshNet Large HCP Atlas	18 subvolumes/sec	10 mb

Download links are in the Example Segmentation Notebooks

Example Segmentation Notebooks

• Gray White Matter Prediction and Visualization
• Mindboggle Prediction and Visualization
• Neuro UNet/Meshnet training tutorial

Training MeshNet on Mindboggle

You can reproduce MeshNet for Mindboggle with 5 simple steps

• Install requirements

  conda env create -f neuro_conda.yml
  conda activate neuro
  pip install -r ./requirements/requirements.txt
  

• Download data. You need to have an account on the open science framework.

  mkdir Mindboggle_data
  osf -p 9ahyp clone Mindboggle_data
  cp -r Mindboggle_data/osfstorage/Mindboggle101_volumes/ data/Mindboggle_data/
  
  If you don't want to make an OSF account the files can also be downloaded
  the correct folder using these commands:
  
  download-gdrive 1l3YCRW7ezV9pw0e3aeYDFi4VsAm_DO7l data/Mindboggle_data/
  download-gdrive 1fxvqrs98F1Gnozg-HsxCjfTrw5zJbQxm data/Mindboggle_data/
  download-gdrive 1w0VXG7mLkE9tULPIxhnNK0qyguuQ7Zx8 data/Mindboggle_data/
  download-gdrive 1soDBWB0iXb3Dc1XarSKgfgwiKDInSLQw data/Mindboggle_data/
  download-gdrive 1jmwE69imokKJJaPKdCNK0Qlxb_r1bGN4 data/Mindboggle_data/
  
  

• Unzip data and remove archive files

  find data/Mindboggle_101 -name '*.tar.gz'| xargs -i tar zxvf {} -C data/Mindboggle_101
  find data/Mindboggle_101 -name '*.tar.gz'| xargs -i rm {}
  

• Prepare data (31 is the number of classes in the example. 102 is the full number of classes in Mindboggle)

  python neuro/scripts/prepare_data.py ./data/Mindboggle_101 31
  

• Start training

  python neuro/minimal_example.py
  

MRI Segmentation Datasets

Mindboggle dataset uses manual annotations which are considered the gold standard. Manually labeling a single MRI can take a week of expert labeling. The labeling is done using a 2D display, one slice at a time which can lead to accuracy/ consistency issues. For our pre-trained models, we use automated labels from the Human Connectome Project (HCP), which can be downloaded here. The automated labeling tool used (FreeSurfer) employs probabilistic methods with priors to perform segmentation and is the current SOTA. While running FreeSurfer involves more than segmentation, it can take hours to segment a single MRI from FreeSurfer vs. minutes for a MeshNet model.

HCP Data preparation

• Prepare T1 or T2 input with mri_convert from FreeSurfer to conform T1 to 1mm voxel size in coronal slice direction with side length 256. You can skip this step if your T1 image is already with slice thickness 1mm x 1mm x 1mm and 256 x 256 x 256.

mri_convert *brainDir*/t1.nii *brainDir*/T1.nii.gz -c

• Prepare 104 atlas labels from aparc+aseg.nii.gz for HCP atlas segmentation using:

python neuro/scripts/prepare_atlas_data.py --brains_list *brains_list.txt*

• Or Prepare GMWM labels from aparc+aseg.mgz using:

./bin/mk_gwmwm_labels.sh [input_directory] [output_directory]

---