code/README.md

AbdCTBench Code

Training and evaluation pipeline for comorbidity prediction from abdominal CT scans.

What This Code Does

• Train single-task or multi-task models from CSV + PNG data
• Evaluate checkpoints with reproducible metrics
• Load released checkpoints in .safetensors or .pth format

Install

pip install -r requirements.txt

Data Layout

Pass --data_dir pointing to:

data_dir/
├── train.csv
├── val.csv
├── test.csv
└── data/
    ├── <FILE>.png
    └── ...

FILE values in CSV must match PNG basenames.

Train

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

Reproducibility flags:

• --seed 42
• --deterministic

Evaluate

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

Useful options:

• --save_predictions
• --save_metrics
• --only_pred

Checkpoint Folder Contents

Each released model folder contains:

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

DICOM to PNG Pipeline

The DICOM -> STL -> PNG conversion scripts are in:

• dicom_stl_png_pipeline/

Key files:

• dicom2stl.py: converts a DICOM series (folder/zip) to STL
• stl2png_centered.py: renders STL to a centered PNG
• parseargs.py and utils/: argument parsing and volume/mesh helper utilities
• requirements.txt in this folder: extra dependencies for this conversion pipeline

Install pipeline dependencies:

pip install -r dicom_stl_png_pipeline/requirements.txt

Minimal usage:

python dicom_stl_png_pipeline/dicom2stl.py \
  --type skin \
  --output ./sample.stl \
  /path/to/dicom_series_folder

python dicom_stl_png_pipeline/stl2png_centered.py \
  ./sample.stl \
  ./sample.png

Notes:

• This conversion pipeline is optional and separate from model train/test.
• train.py and test.py consume PNG + CSV data and do not run DICOM conversion internally.

Biomarker Config Templates

• config/biomarker_config_single_task_example.yaml
• config/biomarker_config_multitask_example.yaml