Deep Learning Models

Deep Learning Models#

nnUNET#

If you want to use nnU-Net to train your model, please use NeuroPoly’s nnU-Net’s fork (also available on PyPI), which contains custom trainers, inference, and other improvements, and most importantly, which ensures compatibility with SCT (consistent pytorch version between SCT and nnU-Net, more details here). Also check our nnU-Net quickstart guide.

Naming models#

Each model is saved in its own repository under the ivadomed organization. The convention for naming repositories is the following:

Syntax:

model_task_anatomy_animal_pathology_condition_contrast_architecture

Keys/Values:

- task = {seg, label, find}, default=None
- anatomy = {sc, gm, csf, brainstem, axon, myelin, ...}, default=sc
- animal = {human, dog, cat, rat, mouse, ...}, default=human
- pathology = {hc, ms, sci}, default=hc
- condition = {invivo, exvivo}, default=invivo
- contrast = {t1, t2, t2star, dwi, sem, tem, oi, ...}, default=None
- architecture = {unet2d, unet3d, filmCharley, hemisAndreanne}, default=unet2d

[!NOTE]
Default fields can be ommited to avoid long names.

Examples:

model_seg_monkey_sc_t1_unet3d
model_seg_sc-gm-lesion_human_ms_exvivo_t2star
model_seg_sc-gm_t1-t2_unet3d  # multi-channel, multi-class

Packaging models#

Models to be used by 3rd party software (e.g. SCT) should be uploaded as ‘assets’ to a release of the repository. The steps are:

  1. Create a release of the repository. The tag and title of the release should be rYYYYMMDD, example: r20240915.

  2. Prepare a folder with the model in the following format. Keep the nnUNet folder structure intact. To save space, include only checkpoint_final.pth. Example:

$ tree model-spinal-rootlets-mp2rage-r20240915.zip
└── nnUNetTrainer_2000epochs__nnUNetPlans__3d_fullres
    ├── dataset.json
    ├── dataset_fingerprint.json
    ├── fold_0
    │         ├── checkpoint_final.pth
    │         ├── debug.json
    │         ├── progress.png
    │         ├── training_log_2024_9_3_14_06_03.png
    │         └── training_log_2024_9_3_14_06_03.txt
    └── plans.json

Important

The dataset.json file should contain the image_orientation entry, for example: "image_orientation": "RPI". This entry is used by SCT to determine the image orientation. But, it is also a good idea to document which orientation was used during training (for the sake of reproducibility).

  1. Zip the folder and upload it as an asset in the release

  2. Publish the release.

Page Contents