Official repository for contrast-agnostic spinal cord segmentation project using SoftSeg.
This repo contains all the code for data preprocessing, training and running inference on other datasets. The code is mainly based on Spinal Cord Toolbox and MONAI (PyTorch).
CITATION INFO: If you find this work and/or code useful for your research, please cite our paper:
@article{bedard2023towards,
title={Towards contrast-agnostic soft segmentation of the spinal cord},
author={B{\'e}dard, Sandrine and Enamundram, Naga Karthik and Tsagkas, Charidimos and Pravat{\`a}, Emanuele and Granziera, Cristina and Smith, Andrew and Weber II, Kenneth Arnold and Cohen-Adad, Julien},
journal={arXiv preprint arXiv:2310.15402},
year={2023}
url={https://arxiv.org/abs/2310.15402}
}
- 1. Main Dependencies
- 2. Dataset
- 3. Preprocessing
- 4. Training
- 5. Computing morphometric measures (CSA)
- 6. Analyse CSA and QC reports
- 7. Get QC reports for other datasets
- 8. Active learning procedure
- SCT 6.2
- Python 3.9
The source data can be found at spine-generic multi-subject.
The preprocessed data are located at duke:projects/ivadomed/contrast-agnostic-seg/data_processed_sg_2023-08-08_NO_CROP\data_processed_clean (internal server)
Main preprocessing steps include:
For T1w and T2w:
- Resampling
- Reorientation to RPI
- Remove suffix
_RPI_r
For T2star:
- Compute root-mean square across 4th dimension (if it exists)
- Remove suffix
_rms
For DWI:
- Generate mean image after motion correction
- Remove intermediate files
Next steps are to generate the contrast-agnostic soft segmentation:
- creates spinal cord segmentation (if it doesn't exist yet)
- creates a mask around the T2 spinal cord
- co-register all contrasts to the T2 spinal cord
- average all segmentations from each contrast within the same space (the T2)
- bring back the segmentations to the original image space of each contrast (except for the T2)
The output of this script is a new derivatives/labels_softseg/ folder that contains the soft labels to be used in this contrast-agnostic segmentation project. All the registration were manually QC-ed (see Quality Control) and the problematic registrations were listed in exclude.yml. The processing was run again to generate the soft segmentations.
Specify the path of preprocessed dataset with the flag -path-data.
This section assumes that SCT is installed. The installation instructions can be found here.
cd processing_spine_generic
sct_run_batch -jobs -1 -path-data <PATH_DATA> -path-output <PATH-OUTPUT> -script process_data.sh -script-args exclude.yml
or use a config file:
Example config_process_data.json:
{
"path_data" : "~/data_nvme_sebeda/datasets/data-multi-subject/",
"path_output" : "~/data_nvme_sebeda/data_processed_sg_2023-08-04_NO_CROP",
"script" : "process_data.sh",
"jobs" : 50,
"exclude_list": ["sub-brnoUhb02", "sub-brnoUhb03", "sub-brnoUhb07", "sub-brnoUhb08", "sub-brnoUhb08", "sub-brnoUhb08", "sub-ucdavis01", "sub-ucdavis02", "sub-ucdavis03", "sub-ucdavis04", "sub-ucdavis05", "sub-ucdavis06", "sub-ucdavis07", "sub-beijingVerio01", "sub-beijingVerio02", "sub-beijingVerio03", "sub-beijingVerio04", "sub-beijingGE01", "sub-beijingGE02", "sub-beijingGE03", "sub-beijingGE04", "sub-ubc01", "sub-oxfordOhba02"]
}
sct_run_batch -config config_process_data.json
A process_data_clean folder is created in where the cropped data and derivatives are included. Here, only the images that have a manual segmentation and soft segmentation are transfered.
After running the analysis, check your Quality Control (QC) report by opening the file /qc/index.html. Use the "search" feature of the QC report to quickly jump to segmentations or labeling issues.
1. Segmentations
If segmentation issues are noticed while checking the quality report, proceed to manual correction using the procedure below:
- In QC report, search for "deepseg" to only display results of spinal cord segmentation
- Review segmentation and spinal cord.
- Click on the F key to indicate if the segmentation is OK ✅, needs manual correction ❌ or if the data is not usable
⚠️ (artifact). Two .yml lists, one for manual corrections and one for the unusable data, will automatically be generated. - Download the lists by clicking on Download QC Fails* and on Download QC Artifacts.
Proceed to manual correction using FSLeyes or ITK snap. Upload the manual segmentations (_seg-manual.nii.gz) with json sidecar in the derivatives. Re-run the analysis: Launch processing
2. Registrations
- In QC report, search for "sct_register_multimodal" to only display results of registration.
- Click on the F key to indicate if the registration is OK ✅, needs to be excluded ❌ or if the data is not usable
⚠️ (artifact). Two .yml lists, will automatically be generated. - Download the list by clicking on Download QC Fails and add the file names under
FILES_REGto https://github.com/sct-pipeline/contrast-agnostic-softseg-spinalcord/blob/main/exclude.yml - Re-run the analysis: Launch processing
3. Soft segmentations
- In QC report, search for "softseg" to only display results of spinal cord soft segmentation
- Click on the F key to indicate if the soft segmentation is OK ✅, needs manual correction ❌ or if the data is not usable
⚠️ (artifact). Two .yml lists, one for manual corrections and one for the unusable data, will automatically be generated. - Download the list by clicking on Download QC Fails.
- Upload the soft segmentations under
data-multi-subject/derivatives/labels_softseg. - Re-run the analysis: Launch processing
The following commands show how to set up the environment. Note that the documentation assumes that the user has conda installed on their system. Instructions on installing conda can be found here.
- Create a conda environment with the following command:
conda create -n venv_monai python=3.9
- Activate the environment with the following command:
conda activate ven_monai
- Install the required packages with the following command:
pip install -r monai/requirements.txt
The training script expects a datalist file in the Medical Decathlon format containing image-label pairs. The datalist can be created by running the create_msd_data.py script. For example, creating the datalist for the soft_all model:
python monai/create_msd_data.py -pd ~/duke/projects/ivadomed/contrast-agnostic-seg/data_processed_sg_2023-08-08_NO_CROP\data_processed_clean> -po ~/datasets/contrast-agnostic/ --contrast all --label-type soft --seed 42
The dataset split containing the training, validation, and test subjects can be found in the monai/data_split_all_soft_seed15.yaml file.
Note The output of the above command is just
.jsonfile pointing to the image-label pairs in the original BIDS dataset. It does not copy the existing data to the output folder.
The training uses MONAI functions and is written in PyTorch Lightning. Example training command to run the soft_all model:
python monai/main.py -m nnunet -crop 64x192x320 --contrast all --label-type soft -initf 32 -me 200 -bs 2 -opt adam -lr 1e-3 -cve 5 -pat 20 -epb -stp --enable_DS
Example training command to run the soft_per_contrast model on the dwi contrast:
python monai/main.py -m nnunet -crop 64x192x320 --contrast dwi --label-type soft -initf 32 -me 3 -bs 2 -opt adam -lr 1e-3 -cve 5 -pat 20 -epb -stp --enable_DS
These commands assume that the datalist created in Section 4.2 lies in the same folder as monai/main.py. Run python monai/main.py -h to see all the available arguments and their descriptions.
Note WandB is used experiment tracking and is implemented via Lightning's Wandblogger. Make sure that the
projectandentityare changed to the appropriate values.
Inference can be run on single images using the monai/run_inference_single_image.py script. Run monai/run_inference_single_image.py -h for usage instructions. Both CPU and GPU-based inference are supported.
To compute the ANIMA metrics shown in the paper, the scripts compute_anima_metrics_*.py are used. For generating the metrics on the spine-generic dataset and also for deepseg and propseg methods, use the following command:
python anima_metrics/compute_anima_metrics_spine_generic.py --pred-folder <PATH_PREDS> --method <monai/deepseg2d/deepseg3d/propseg> -dname spine-generic
For reproducing the results on the other datasets, use the following command:
python anima_metrics/compute_anima_metrics_unseen_datasets.py --pred-folder <PATH_PREDS> -dname <sci-t2w/ms-mp2rage/radiculopathy-epi>
Note The
--pred-folderargument expects the path to the folder containing the prediction and GT segmentation masks.
To compute the CSA at C2-C3 vertebral levels on the prediction masks and get the QC report of the predictions, the script compute_csa_qc_<nnunet/monai>.sh are used. The input is the folder data_processed_clean (result from preprocessing) and the path of the prediction masks is added as an extra script argument -script-args.
For every trained model, you can run:
sct_run_batch -jobs -1 -path-data /data_processed_clean/ -path-output <PATH_OUTPUT> -script compute_csa_qc_<nnunet/monai>.sh -script-args <PATH_PRED_MASKS>
-path-data: Path to data from spine generic used for training.-path-output: Path to save results-script: Script to compute the CSA and QC report-script-args: Path to the prediction masks
The CSA results will be under <PATH_OUTPUT>/results and the QC report under <PATH_OUTPUT>/qc.
Here is an example on how to compute CSA and QC on contrast-agnostic model
sct_run_batch -jobs -1 -path-data ~/duke/projects/ivadomed/contrast-agnostic-seg/data_processed_sg_2023-03-10_NO_CROP\data_processed_clean -path-output ~/results -script compute_csa_qc_monai.sh -script-args ~/duke/projects/ivadomed/contrast-agnostic-seg/models/monai/spine-generic-results
Note: For nnUnet, change the variable prefix in the script compute_csa_nnunet.sh according to the prefix in the prediction name.
Here is an example on how to compute CSA and QC on nnUNet models.
sct_run_batch -jobs -1 -path-data ~/duke/projects/ivadomed/contrast-agnostic-seg/data_processed_sg_2023-03-10_NO_CROP\data_processed_clean -path-output ~/results -script compute_csa_qc_nnunet.sh -script-args ~/duke/projects/ivadomed/contrast-agnostic-seg/models/nnunet/spine-generic-results/test_predictions_2023-08-24
To generate violin plots and analyse results, put all CSA results file in the same folder (here csa_ivadomed_vs_nnunet_vs_monai) and run:
python analyse_csa_all_models.py -i-folder ~/duke/projects/ivadomed/contrast-agnostic-seg/csa_measures_pred/csa_ivadomed_vs_nnunet_vs_monai/ \
-include csa_monai_nnunet_2023-09-18 csa_monai_nnunet_per_contrast csa_gt_2023-08-08 csa_gt_hard_2023-08-08 \
csa_nnunet_2023-08-24 csa_other_methods_2023-09-21-all csa_monai_nnunet_2023-09-18_hard csa_monai_nnunet_diceL
-i-folder: Path to folder containing CSA results from models to analyse-include: names of the folder names to include in the analysis (one model = one folder)
The plots will be saved to the parent directory with the name charts_<datetime.now())>
The QC reports from three other datasets sci-t2w, ms-mp2rage, and radiculopathy-epi are shown in the paper. The scripts for reproducing the results are in the qc_other_datasets folder.
General command to run QC on prediction masks from other datasets:
sct_run_batch -path-data <PATH_DATA> -path-out <PATH-OUT> -script-args <PATH_PRED_MASK> -jobs 20 -script run_qc_prediction_<dataset>.sh
-path-data: Path to the original dataset used to run inferences.-path-output: Path to the results folder to save QC report-script: Scriptrun_qc_prediction_<dataset>corresponding to the dataset.-script-args: Path to prediction masks for the specific dataset
Using the contrast-agnostic model:
sct_run_batch -jobs 32 -path-data ~/path-to-dataset/sci-colorado/ \
-path-output ~/path-to-output/qc_contrast-agnostic_sci-colorado \
-script run_qc_prediction_sci_colorado.sh \
-script-args ~/duke/projects/ivadomed/contrast-agnostic-seg/models/monai/sci-colorado-results/test_preds_colorado_soft_all
Using the nnUNet model:
sct_run_batch -jobs 32 -path-data ~/path-to-dataset/sci-colorado/ \
-path-output ~/path-to-output/qc_nnunet_sci-colorado \
-script run_qc_prediction_sci_colorado.sh \
-script-args "/home/GRAMES.POLYMTL.CA/u114716/duke/projects/ivadomed/contrast-agnostic-seg/models/nnunet/sci-colorado-results/test_predictions nnUNet"
Using the contrast-agnostic model:
sct_run_batch -jobs 32 -path-data ~/path-to-dataset/basel-mp2rage/ \
-path-output ~/path-to-output/qc_contrast-agnostic_basel-mp2rage \
-script run_qc_prediction_basel_mp2rage.sh \
-script-args ~/duke/projects/ivadomed/contrast-agnostic-seg/models/monai/basel-mp2rage-rpi-results/test_preds_mp2rage_soft_all
Using the nnUNet model:
sct_run_batch -jobs 32 -path-data ~/path-to-dataset/basel-mp2rage/ \
-path-output ~/path-to-output/qc_nnunet_basel-mp2rage \
-script run_qc_prediction_basel_mp2rage.sh \
-script-args "/home/GRAMES.POLYMTL.CA/u114716/duke/projects/ivadomed/contrast-agnostic-seg/models/nnunet/basel-mp2rage-rpi-results/test_predictions nnUNet"
Using the contrast-agnostic model:
sct_run_batch -jobs 32 -path-data ~/path-to-dataset/epi-stanford/ \
-path-output ~/path-to-output/qc_contrast-agnostic_epi-stanford \
-script run_qc_prediction_epi_stanford.sh \
-script-args ~/duke/projects/ivadomed/contrast-agnostic-seg/models/monai/epi-stanford-results/test_preds_soft_all
Using the nnUNet model:
sct_run_batch -jobs 32 -path-data ~/path-to-dataset/epi-stanford/ \
-path-output ~/path-to-output/qc_nnunet_radiculopathy-epi \
-script run_qc_prediction_epi_stanford.sh \
-script-args "/home/GRAMES.POLYMTL.CA/u114716/duke/projects/ivadomed/contrast-agnostic-seg/models/nnunet/epi-stanford-results/test_preds_stanford_rest_weber nnUNet"
Note This section is still a work in progress.
To extend the training set to other contrasts and to pathologies, we applied the segmentation model to other datasets, manually corrected the segmentations and added them to the training set.
Here is the detailed procedure:
- Run inference on other datasets for the selected models and generate the QC report from prediction masks.
- Select ~20 interesting images per dataset (using the QC report).
- Correct the inference on the selected subjects if needed (you can use
manual-correctionscript). - Add the inferred segmentations to the
derivatives/labels_contrast_agnosticfolder of each dataset. - Add inferred segmentations to the training set (keep the same testing spine generic subjects) & retrain a model.
- Compute CSA on spine generic testing set and see STD vs before