contrast-agnostic-spinal-cord-v3.0 (r20250402)

Overview of the dataset characteristics

**List of datasets used **

• basel-mp2rage
• canproco
• data-multi-subject
• dcm-brno
• dcm-zurich-lesions-20231115
• dcm-zurich-lesions
• dcm-zurich
• lumbar-epfl
• lumbar-vanderbilt
• nih-ms-mp2rage
• sci-colorado
• sci-paris
• sci-zurich
• sct-testing-large (T2star and MTon contrasts)
• site_006 (praxis - montreal site)
• site_007 (praxis - vancouver site)

** Dataset stats I (SUBJECT-WISE PATHOLOGY SPLIT) **

Pathology	Number of Subjects
ALS	13
AcuteSCI	95
DCM	359
HC	428
MS	164
NMO	10
PPMS	60
RIS	61
RRMS	249
SCI	191
SYR	1
TOTAL	1631

** Dataset stats II (CONTRAST-WISE PATHOLOGY SPLIT) **

	MS	HC	RRMS	RIS	PPMS	DCM	SCI	ALS	NMO	SYR	AcuteSCI	#total_per_contrast
dwi	0	184	0	0	0	59	0	0	0	0	0	243
mt-off	0	184	0	0	0	59	0	0	0	0	0	243
mt-on	0	184	0	0	0	64	0	0	0	0	0	248
psir	0	42	193	54	44	0	0	0	0	0	0	333
stir	0	10	56	7	16	0	0	0	0	0	0	89
t1w	0	249	0	0	0	59	0	0	10	0	0	318
t2star	121	237	0	0	0	127	0	13	0	1	0	499
t2w	0	252	229	61	57	426	257	0	0	0	95	1377
unit1	50	53	0	0	0	0	0	0	0	0	0	103
TOTAL	171	1395	478	122	117	794	257	13	10	1	95	3453

NOTE: All the info presented above can also be found in the release assets

What’s Changed

• Change in Training Framework: Replacing monai-based model with a 3D nnUNet model trained from scratch. Tested on a wide variety of pathologies and contrasts. Works especially well on compressed spinal cords, fixes issues with shifted predictions (consistent in monai-based models)
• Introducing Lifelong Learning Framework: Adds a GitHub Actions-based workflow to automatically generate plots measuring CSA variability once a release has been published. This addition makes it easy to monitor morphometric drift between different model releases. Currently, the plots are generated only for v2.0 (original model) and v3.0 (the current release)
• Other:
  • Update model with new datasets, pathologies and contrasts by @naga-karthik in #125
  • Lifelong learning strategy for adding new contrasts and pathologies by @naga-karthik in #128
  • Add instructions to use contrast agnostic model with nnUnet by @naga-karthik in #129
  • Update dataset info in README by @sandrinebedard in #135
  • Release new nnunet model replacing the current monai-based model by @naga-karthik in #136
  • Add standalone script to compute morphometrics after training by @naga-karthik in #139
  • Clean repo and provide instructions for future reproducibility and retraining by @naga-karthik in #138
  • update readme by @naga-karthik in #144

Full Changelog: v2.5...v3.0

r20241024

Improved version of contrast-agnostic spinal cord segmentation model trained on healthy subjects and pathologies in the cervical cord. The dataset_stats_overall.txt contains the list of contrasts, pathologies, and the respective splits for each.

Works well on:

• Spinal cord injury (SCI) lesions
• GRE-EPI images
• B0 Field Map images
• Lumbar cord
• PSIR and STIR contrasts
• [NEW] whole-spine images (tested on T1w/T2w only)

What's Changed

• Update preprocessing script for spine-generic with new naming convention by @sandrinebedard in #105
• Continual training of contrast-agnostic model with new contrasts and pathologies by @naga-karthik in #104

Full Changelog: v2.3...v2.4.1-beta

r20240531

Improved version of contrast-agnostic spinal cord segmentation model trained on healthy subjects and pathologies in the cervical cord:

• Contrasts: T1w, T2w, T2star, MTon-MTS, MToff-MTS, DWI (averaged), mp2rage UNIT1, PSIR, STIR
• Pathologies: multiple sclerosis (MS) patients, compressed spinal cords in degenerative cervical myelopathy (DCM) patients.

Works well on:

• Spinal cord injury (SCI) lesions
• GRE-EPI images
• B0 Field Map images
• Lumbar cord
• PSIR and STIR contrasts

Main difference from version v2.3 is the addition of lumbar T2w images and PSIR/STIR contrasts of MS patients.

The train/val/test splits from all the datasets used to train this model can be found in the datasplits folder in the source code. Further details on the number of training samples across all datasets and samples per contrast can be found in dataset_splits.md.

EDIT: the initial .zip file containing the model was corrupted, hence a new (fixed) .zip was uploaded

What's Changed

• Update preprocessing script for spine-generic with new naming convention by @sandrinebedard in #105

Full Changelog: v2.3...v2.4

r20240417

Improved version of contrast-agnostic spinal cord segmentation model trained on healthy subjects and pathologies in the cervical cord:

• Contrasts: T1w, T2w, T2star, MTon-MTS, MToff-MTS, DWI (averaged), mp2rage UNIT1
• Pathologies: multiple sclerosis (MS) patients, compressed spinal cords in degenerative cervical myelopathy (DCM) patients.

Works well on:

• Spinal cord injury (SCI) lesions
• GRE-EPI images
• B0 Field Map images
• Lumbar cord

Main difference from earlier versions of contrast-agnostic is the addition of pathological data (MS, DCM) to the training set.

The train/val/test splits from all the datasets used to train this model can be found in the datasplits folder in the source code. Further details on the number of training samples across all datasets and samples per contrast can be found in dataset_stats.md.

Full Changelog: v2.2...v2.3

r20240529

Official code for the spin-off extension of contrast-agnostic spinal cord segmentation comparing different DL architectures including CNNs, Vision Transformers and ConvNeXT models accepted as a short paper at MIDL 2024. The paper can be found here.

r20240328

What's Changed

• Update preprocessing script for spine-generic with new naming convention by @sandrinebedard in #105

Full Changelog: v2.1...v2.2

r20240307

About

This release updates the training codebase and adds a newer variant of the contrast-agnostic model (details below).

What's Changed

• Update readme with citation info and arxiv badge by @naga-karthik in #88
• Fixed documentation on inference by @jcohenadad in #91
• Fixed documentation on inference (follow-up) by @jcohenadad in #93
• Make inference CLI callable from anywhere by @jcohenadad in #95
• add info about dataset split used for training contrast-agnostic model by @naga-karthik in #97
• Simplify inference script by @naga-karthik in #100
• Port towards config file-based training by @naga-karthik in #90

Other notable changes

• The model in this release is trained with binarized soft labels (hence the name soft_bin) as opposed to directly training on soft labels as in the model in release v2.0
• In addition to the monai-based nnunet model, this release also adds the feature to train other models as well (e.g. SwinUNETR, MedNeXT, etc.)
• Three new classes of CSA evaluation scripts are added -- (1) evaluating CSA across different models, (2) evaluating CSA across different resolutions, and (3) evaluating CSA across different resolutions.
  • A unified script analyze_csa_across.py is added for generating CSA violin plots across different classes mentioned above.

Full Changelog: v2.0...v2.1

contrast-agnostic-softseg-spinalcord v2.0

About

This release contains the official code for the submission to Medical Image Analysis Journal. The model weights are uploaded as release assets along with all the scripts for preprocessing, training, CSA and QC generation.

What's Changed

• Add subjects with missing contrasts to exclude by @sandrinebedard in #53
• Add script to compute CSA binarized and run QC on binarized prediction for ivadomed by @sandrinebedard in #58
• Modify post processing nnunet for CSA computation and QC report by @sandrinebedard in #57
• Naga/nnunet by @sandrinebedard in #65
• add monai-based scripts for dataset conversion, training, and inference by @naga-karthik in #60
• Clean up processing and repository by @sandrinebedard in #64
• add requirements for gpu inference by @naga-karthik in #76
• add qc generation script for epi data by @naga-karthik in #77
• Update main readme by @naga-karthik in #78
• Upgrade and pin monai to 1.3.0 by @naga-karthik in #79
• Finalise script for figures by @sandrinebedard in #80
• add anima metrics by @naga-karthik in #81
• Add anova test on soft all and hard all by @sandrinebedard in #82

New Contributors

• @naga-karthik made their first contribution in #60

Full Changelog: v1.2...v2.0

contrast-agnostic-softseg-spinalcord v1.2 - MICCAI 2023

This code was used for the submission to MICCAI 2023. QC reports of the tested models on Basel-MP2RAGE and sci-colorado datasets are includes as release assets.

What's Changed

• use _pred, not _pred_painted & remove 4th dimension by @sandrinebedard in #35
• CSA charts preliminary notebook by @ebadrian in #30
• Add command for CSA charts by @sandrinebedard in #43
• Change file names for MTS contrasts by @sandrinebedard in #42
• Fix MTS filenames in preprocessing and charts scripts by @sandrinebedard in #46
• Add processing of datasets with pathology by @sandrinebedard in #50
• Training and evaluation scripts MICCAI 2023 by @ebadrian in #51

New Contributors

• @ebadrian made their first contribution in #30

Full Changelog: v1.1...v1.2

contrast-agnostic-softseg-spinalcord v1.1

What's Changed

• The registration algorithm was modified for the soft segmentation generation in the dataset spine-generic/data-multi-subject.