Coronary Artery Segmentation with SegMamba

Resume-ready project for 3D coronary artery segmentation on ASOCA CTCA volumes using the exact SegMamba architecture from the original paper and official codebase.

What This Repository Contains

• Reproducible training entrypoint with CLI config (main.py)
• Sliding-window inference script (inference.py)
• Lightweight volume visualization utility (visualize.py)
• SegMamba model implementation (model_segmamba/segmamba.py)
• Legacy experiment scripts kept for traceability (legacy/)
• Updated short report assets (docs/)

Quick Start

1) Install dependencies (uv)

uv sync

Alternative:

uv venv
source .venv/bin/activate
uv pip install -r requirements.txt

2) Expected dataset layout (ASOCA)

<DATA_ROOT>/
  Diseased/
    CTCA/Diseased_1.nrrd ... Diseased_19.nrrd
    Annotations/Diseased_1.nrrd ... Diseased_19.nrrd
  Normal/
    CTCA/Normal_1.nrrd ... Normal_19.nrrd
    Annotations/Normal_1.nrrd ... Normal_19.nrrd

3) Train

uv run python main.py \
  --data-root /path/to/asoca \
  --output-dir outputs/segmamba_run \
  --epochs 120 \
  --patch-size 224,224,96 \
  --samples-per-volume 3 \
  --val-interval 5 \
  --amp

4) Inference

uv run python inference.py \
  --checkpoint outputs/segmamba_run/checkpoints/best_model.pt \
  --image /path/to/Diseased_1.nrrd \
  --output outputs/segmamba_run/inference_mask.nrrd \
  --patch-size 224,224,96

5) Visualize slices

uv run python visualize.py --volume outputs/segmamba_run/inference_mask.nrrd

Historical Results (from prior report)

These are validation-set numbers from the earlier project phase:

Model	Mean Validation Dice
UNet baseline	0.7844
Mamba-Encoder	0.7734
SegMamba	0.7673

See the report in docs/report.pdf.

Reports

• Report PDF: docs/report.pdf
• LaTeX source: docs/report.tex

Notes

• legacy/ contains older scripts retained for experiment traceability.
• Checkpoints and generated artifacts are ignored via .gitignore.