COMPASS: Robust Feature Conformal Prediction for Medical Segmentation Metrics

This repository contains the COMPASS 2D code for conformal prediction on segmentation-derived scalar metrics.

Overview

Citation

If you use COMPASS in your research, please cite:

@article{cheung2025compass,
  title={COMPASS: Robust Feature Conformal Prediction for Medical Segmentation Metrics},
  author={Cheung, Matt Y and Veeraraghavan, Ashok and Balakrishnan, Guha},
  journal={arXiv preprint arXiv:2509.22240},
  year={2025}
}

Quickstart (2D)

1. Install:

pip install -r requirements-2d.txt

2. Prepare dataset folders and train models:

• preprocessing + training helpers: 2D/datasets/<dataset>/
• baseline training: python 2D/datasets/<dataset>/run_nn.py
• 3-head QR/CQR training: python 2D/datasets/<dataset>/run_qrnn.py

3. Run COMPASS:

• recommended (logits-only): python 2D/run_cp_L.py
• full (includes Jacobians): python 2D/run_cp.py

See 2D/README.md for step-by-step details.

Results

Outputs are written under COMPASS_RESULTS_DIR (default: /scratch/yc130/compass_results).

This release intentionally does not include figure-generation / plotting scripts.