Reproducible cardiac MRI SAX segmentation workflows with public ACDC validation, Dice/HD95 reporting, QC overlays, and failure-mode analysis.
Public ACDC fold 0 validation, nnU-Net v2 2D CPU debug run, conservative anatomical postprocessing.
| ACDC fold 0 | RV | Myocardium | LV |
|---|---|---|---|
| Dice | 0.8579 | 0.8547 | 0.9093 |
| HD95 | 7.44 | 4.49 | 3.99 |
Mean Dice: 0.8740; mean HD95: 5.31.
This is a reproducibility milestone, not a clinical model or state-of-the-art claim. The next performance milestone is full GPU nnU-Net training with broader validation.
| Area | Included artifacts |
|---|---|
| Data preparation | ACDC, MSD Cardiac debug, local CorSeg pseudo-label converters |
| Training workflow | nnU-Net shell wrappers and local debug trainer installer |
| Evaluation | Dice/HD95 CSV and Markdown reports |
| QC visualization | Expert-label, prediction, and failure-case overlays |
| Research hygiene | Privacy statement, release checklist, launch notes, starter issues |
The repository does not include private patient data, DICOM files, NIfTI files, trained checkpoints, or private source manifests.
Install the lightweight project dependencies:
pip install -r requirements.txt
python scripts/install_tiny_debug_trainer.pyPrepare public ACDC data in nnU-Net format:
python scripts/prepare_acdc.py \
--source /path/to/acdc_raw \
--output "$nnUNet_raw/Dataset001_ACDC"Train, predict, evaluate, and render QC:
DATASET_ID=1 CONFIG=2d FOLD=0 DEVICE=cuda bash scripts/train_nnunet_acdc.sh
DATASET_ID=1 CONFIG=2d FOLD=0 bash scripts/predict_nnunet_acdc.sh
python scripts/evaluate_segmentation.py \
--pred-dir predictions/acdc_nnunet_2d_fold0 \
--label-dir "$nnUNet_raw/Dataset001_ACDC/labelsTr" \
--out-csv results/acdc_metrics.csv \
--out-md results/acdc_metrics.md
python scripts/visualize_segmentation.py \
--image "$nnUNet_raw/Dataset001_ACDC/imagesTr/patient001_frame01_0000.nii.gz" \
--label "$nnUNet_raw/Dataset001_ACDC/labelsTr/patient001_frame01.nii.gz" \
--prediction predictions/acdc_nnunet_2d_fold0/patient001_frame01.nii.gz \
--output results/figures/acdc_patient001_frame01_overlay.pngFor a fuller setup walkthrough, see docs/quickstart.md.
Representative validation prediction:
Representative failure case:
Key result notes:
- ACDC nnU-Net baseline
- Postprocessed fold 0 metrics
- Failure analysis
- Tiny smoke test
- MSD Cardiac debug run
- CorSeg SAX pseudo-label debug run
cmr-ai-segmentation-workbench/
├── configs/ # local environment examples
├── docs/ # data, privacy, launch, and troubleshooting docs
├── docs/assets/ # derived PNG/SVG figures for README and sharing
├── results/ # metrics and result notes, no raw medical data
├── scripts/ # data preparation, evaluation, visualization, wrappers
├── CHANGELOG.md
├── CITATION.cff
└── README.md
| Priority | Milestone | Why it matters |
|---|---|---|
| 1 | Full GPU nnU-Net 2D ACDC baseline | Establish a fairer public benchmark |
| 2 | MONAI ACDC baseline | Add an independent implementation path |
| 3 | ED vs ES stratified reporting | Clarify small-cavity and phase-specific failure modes |
| 4 | 3D full-resolution baseline | Test spatial context for basal/apical errors |
| 5 | Paper-ready technical report | Convert workflow evidence into a citable research artifact |
Open tracking issue: Add minimal MONAI baseline. Community feedback request: MONAI discussion #8899.
- Use public datasets for reproducible examples.
- Keep private DICOM, NIfTI, source manifests, and checkpoints outside Git.
- Do not describe pseudo-label debug runs as expert-labeled model performance.
- Do not use this project for diagnosis, treatment decisions, or regulated clinical deployment.
See docs/model_and_data_statement.md, docs/medical_disclaimer.md, and docs/release_checklist.md.
This workbench builds around excellent open-source medical AI projects:


