r20250219

What's Changed

• Dataset aggregation by @plbenveniste in #46
• Add nnUNet training by @plbenveniste in #47
• Dataset analysis of MSD dataset by @plbenveniste in #48
• Evaluation of existing models on the different sets by @plbenveniste in #49

Full Changelog: r20241101...r20250219

Release content:

This release corresponds to the code used for the ACTRIMS/NAIMS poster 2025 “Automatic multi-contrast MRI segmentation of spinal cord lesions”. It contains dataset_2025-01-17_seed42.json which stores the data split used in this project a well as the fold of the model.

The pipeline used for training is the following:

Because of size constraints, the models were stored in multiple files containing each split. To use them, they should be assembled in the following way:

The datasets used are:

• basel-mp2rage: c54b3fb0fad8fa07baedb724ce9d919a73854e6e
• bavaria-quebec-spine-ms-unstitched: f44f92b39eadd4276f0f689c96e63954051aea32
• canproco: 7daf5ed2abc4304b45d6c81364ed06df0fc1f7c4
• ms-basel-2018: cfddb0ed1aee701ef512631edb6016bcc29ef47b
• ms-basel-2020: 92ec31c142bc01cc37957742fcf318d55b7ef80d
• ms-karolinska-2020: d81e0c7a5aa8092df274f8d88827b64b4e1b4dc4
• ms-nmo-beijing: 3359fc242e02367bb2373fd3f8809c1ea8624ced
• ms-nyu: fd9f1880d639e7ef169daa11396c1f1b1687fde9
• ms-rennes-mp2rage: 4a6b2d32bf7dca400d96a17d71eb749b0db4b4eb
• nih-ms-mp2rage: 14b7c9fbc73613ae49cdbacb555a55d8fb071605
• sct-testing-large: 0299da1367ac0958e94d3af39a14a0382f14de00
• umass-ms-ge-excite1.5: 18125775b55a21aac1ee8fcb0cc60cd1622e5349
• umass-ms-ge-hdxt1.5: 503f28b65b0d3cdbb99f790daec4a976b35c0ffc
• umass-ms-ge-pioneer3: 89cf98a82f951c96cd9517cc5e201a36be490dfd
• umass-ms-siemens-espree1.5: ab24170b5a0e1fbbb63d5631d1a7052ad0197982

The results of the poster were computed using the “checkpoint_best.pth” models.