README.md

metadata

license: cc-by-4.0
language:
  - en
size_categories:
  - 10K<n<100K

AbdCTBench: Data + Code + Checkpoints

AbdCTBench is a public benchmark for comorbidity prediction from abdominal CT scans. This repository contains:

• AbdCTBench_dataset/: split CSVs + image data archive + 3D stl meshes
• code/: training and evaluation pipeline
• models/: released pretrained checkpoints (.safetensors) with configs
• code/dicom_stl_png_pipeline/: DICOM -> STL -> PNG conversion scripts

Quick Start

From code/:

pip install -r requirements.txt

Use dataset in this same repo:

cd code
python test.py \
  --data_dir ../AbdCTBench_dataset \
  --checkpoint_path ../models/mi_only/ResNet-18_lr1e-05_bs16/best_checkpoint.safetensors \
  --biomarker_config ../models/mi_only/ResNet-18_lr1e-05_bs16/biomarker_config.json \
  --output_dir ./test_results

Training

cd code
python train.py \
  --model "ResNet-18" \
  --data_dir ../AbdCTBench_dataset \
  --biomarker_config ./config/biomarker_config_multitask_example.yaml \
  --output_dir ./outputs

Reproducibility controls:

• --seed (default 42)
• --deterministic (slower but more deterministic backend behavior)

Checkpoint Format

Each released model folder under models/ contains:

• best_checkpoint.safetensors
• config.json
• biomarker_config.json

test.py supports both .safetensors and .pth checkpoints.

Expected Data Layout

--data_dir should contain:

• train.csv, val.csv, test.csv
• data/ with PNG files named by FILE column values

Citation

If you use AbdCTBench, please cite:

@inproceedings{
chaudhry2026abdctbench,
title={Abd{CTB}ench: Learning Clinical Biomarker Representations from Abdominal Surface Geometry},
author={Muhammad Ahmed Chaudhry and Suhana Bedi and Pola Lydia Lagari and Brian T Layden and William Galanter and Ayis Pyrros and Sanmi Koyejo},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=dKRAo0a9Gm}
}

---

For detailed internals and full argument descriptions, see code/README.md.