Cortical Parcellation tool

This program quickly labels the human brain cortical regions, from a pre-segmented mask input.

The input can be a probabilistic (e.g. SPM c1/wc1 images) or a boolean mask (e.g. FSL and other).

The regions follow one of the Desikan-Killiany (DK), Destrieux 2009, or PALS_Brodman definition, originally defined on the surface.

This is based on the methods described in the manuscript Learning a cortical parcellation of the brain robust to the MRI segmentation with convolutional neural networks ( https://doi.org/10.1016/j.media.2020.101639 ). It contains background information and more.

This version is a direct PyTorch port of the model originally developed and trained on the (now unmaintained) Theano framework. The theano-based inference code is still available in the "theano" git branch of this repository. Both version use the same parameters and should give identical results.

Installation

This code uses pytorch, and relies on ANTs. No GPU is required.

To setup a ANTs environment, get it from http://stnava.github.io/ANTs/ (or alternatively, from a docker container). The 2.1.0 binaries are known to work.

The simplest way to install the rest from scratch is to use a Anaconda or virtualenv environment using Python >3.5, then

• install numpy, scipy, if not installed (conda install scipy or pip install scipy)
• install nibabel, available on pip (pip install nibabel)
• install pytorch ( > 1.0.0, tested until 1.4.0 ) following the instruction of https://pytorch.org (CPU only is ok, no need for CUDA/GPU-related dependencies)

Usage:

After download, you can run

./test_run.sh in the source directory to ensure the environment is ok.

To use the program, simply run:

./parcel_seg.sh -a aseg example/example_segmentation_t1.nii.gz.

For more flexibility, the following options are available:

Options: -c N : extract label N and use that for input, useful if the input is a label map.
         -a X : use atlas X, where X is one of: aseg a2009 apals. Default to all
         -n   : skip the internal registration step, useful if the input is already aligned to MNI space
         -t X : threshold the input segmentation at X instead of 0.5 when creating the final cortical label map.
         -d   : debug - keep all temporary files