README.md

metadata

license: mit
language:
  - en
library_name: pytorch
tags:
  - medical-imaging
  - coronary-artery-segmentation
  - cardiac-ct
  - nnunet
  - imagecas

Coronary Artery Segmentation (nnU-Net)

nnU-Net trained on ImageCAS dataset for coronary artery tree segmentation in cardiac CT angiography (CCTA).

Model Details

• Architecture: nnU-Net v2 (3D Residual-Encoder, full resolution)
• Input: 3D CCTA volumes (NIfTI format)
• Output: Coronary artery voxel-wise segmentation

Training

• Dataset: ImageCAS (~1,000 CCTA scans)
• Split: Official ImageCAS split (train/val/test)
• Fold: fold_0 (best validation performance)

Downstream Pipeline

Part of multi-stage clinical analysis system:

1. Stage 1 (this model): Coronary segmentation
2. Stage 2: Plaque classification (HU-threshold: calcified/non-calcified/LRNC)
3. Stage 3: 3D plaque map reconstruction
4. Stage 4: FFR estimation (research prototype, CFD-capable)
5. Stage 5: Clinical review app with 3D visualization & curved MPR

Real Case Examples

Outputs included demonstrate:

• Coronary mesh with plaque composition color-coding (phase4_job_case1_v2)
• Curved MPR for interventionalist visualization
• Diameter profiling, QC gates, comprehensive reports

Citation

ImageCAS: Huo et al., "Towards Robust Coronary Artery Segmentation in CCTA" (2021)

nnU-Net: Isensee et al., "nnU-Net: Self-configuring method for deep learning-based biomedical image segmentation" (2021)

License

MIT - See repository for details

Disclaimer

Research model. Not for clinical use without independent validation.