mappo23 / dl-direct-v8

DL+DiReCT-v8 - Deep Learning-based Lesion-aware Brain Anatomy Segmentation and Parcellation Model

About DL+DiReCT v8

DL+DiReCT combines a deep learning-based neuroanatomy segmentation and cortex parcellation with a diffeomorphic registration technique to measure cortical thickness from T1w MRI.

This repo is a copy of the original and serves as a demo for the new retrained version of the model (i.e. v8). With this retrained version, the model gains the ability to be aware of lesions and to segment related hypointensity regions.

In /pat_test_data folder we include few patient data examples in order to test model.

To ensure consistency with the companion thesis work, we will cover the segmentation capabilities only of the DL+DiReCT pipeline.

If you are using DL+DiReCT in your research, please cite (bibtex) the corresponding publication:

Rebsamen, M, Rummel, C, Reyes, M, Wiest, R, McKinley, R.
Direct cortical thickness estimation using deep learning‐based anatomy segmentation and cortex parcellation.
Human brain mapping. 2020; 41: 4804-4814. https://doi.org/10.1002/hbm.25159

Installation

Create virtual environment (optional)

Download and install Miniconda and create a new conda environment:

conda create -y -n DL_DiReCT python=3.10
source activate DL_DiReCT

Install DL+DiReCT

cd ${HOME}
git clone https://github.com/Mappo23/DL-DiReCT-v8
cd DL-DiReCT-v8
pip install numpy && pip install -e .

Usage

Run dl+direct on a patient T1-weighted MRI with:

source activate DL_DiReCT

dl+direct -n -k --model v8 <path_to_t1_input.nii.gz> <output_dir>

In environments with limited RAM capabilities (i.e. less than 16GB), we recommend using the flag --lowmem to ensure proper segmentation execution.

Following files of interest are generated in the output directory:

- T1w_norm.nii.gz		Re-sampled input volume
- T1w_norm_seg.nii.gz		Segmentation
- seg_<ROI>.nii.gz              <ROI> label probability map   
- seg_lesion.nii.gz             lesion label probability map   
- result-vol.csv		Segmentation volumes
- label_def.csv			Label definitions of the segmentation

lesion region is identified by label code 50000, the associated color is bright yellow.

To inspect segmentation results (i.e. T1w_norm_seg.nii.gz), it is recommended to utilise specialised software. Here, you can find a quick overview of the main options.

Other Available Models

The following models are available with the --model ... option:

• v0: Default, for un-enhanced T1w MRI, cortex parcellation with Desikan-Killiany atlas (Rebsamen et al., 2020)
• v6: For both contrast-enhanced and un-enhanced MRI (Rebsamen et al., 2022)
• v7: Same as v6, with 74 region per hemisphere according the Destrieux atlas (Rebsamen et al., 2022)

Frequently Asked Questions

For further details, consult the corresponding publication and the FAQ or contact us