lumbar_seg_EPFL

Model repository for lumbar spinal cord segmentation from EPFL data.

Context and objective

Sciatica or hernias in the lumbar spinal cord can cause severe chronic pain. Researchers at EPFL, in Switzerland, shared with our laboratory MRI images of the lumbar marrow of participants, as well as manual segmentations made by experts.

Here, we aim to train a lumbar spinal cord segmentation model on these MRI images and make the model accessible via open-source software to enable global use.

Model overview

Data

This dataset was shared by Nawal Kinany [email protected] and Dimitry Van De Ville [email protected], from EPFL.

This dataset contains images of the lumbar spinal cord with T2w contrast.

Access to the data

(internal) [email protected]:datasets/lumbar-epfl

Getting started

Dependencies

• SCT commit: git-master-7c4f081a0cb566fd8f4702a7e0f8b43bcb02b412
• ivadomed commit: git-master-97b5772374660b12895f7a458941f468090e8bf8

Preprocess the data

The data need to be preprocessed before training. This will reorient the images to a common orientation (LPI) and resample them. To run the script, run the following command

sct_run_batch -script <PATH_TO_REPOSITORY>/preprocess/preprocess_data.sh -path-data <PATH_TO_DATA>/lumbar_epfl/ -path-output <PATH_OUTPUTS>/lumbar_epfl_preprocessed -jobs <JOBS>

Where:

• <JOBS>: Number of CPU cores to use (we recommend not using more than half the number of available cores.)

Training

The training task is carried out using the clean preprocessed dataset using the command

ivadomed --train -c config.json --path-data path/to/bids/data --path-output path/to/output/directory

Testing

Testing a lumbar cord segmentation is done using the command

ivadomed --test -c config.json --path-data path/to/bids/data --path-output path/to/output/directory

Segment a dataset

In order to segment a dataset, run the following command

ivadomed --segment -c config.json --path-data path/to/bids/data --path-output path/to/output/directory