GenEval: Multi-Modal Single-Source Domain Generalization for Medical Imaging

Official PyTorch implementation of GenEval (WACV 2026).

Authors: Ayan Banerjee, Kuntal Thakur Sandeep Gupta
Affiliation: IMPACT Lab, Arizona State University

Overview

GenEval combines multi-modal learning with single-source domain generalization for diabetic retinopathy (DR). It fine-tunes MedGemma-4B with LoRA adapters to classify DR severity grades across multiple retinal datasets.

Installation

git clone https://github.com/ImpactLabASU/GenEval.git
cd GenEval
pip install -r requirements.txt

See docs/DATASETS.md for details.

Training

python diabetic_retinopathy/train.py \
  --dataset aptos \
  --data_path /path/to/aptos/images \
  --csv_path /path/to/aptos/labels.csv \
  --output_dir ./checkpoints/aptos

Multi-domain (MDG-style) example:

python diabetic_retinopathy/train.py \
  --datasets messidor2 eyepacs \
  --data_paths /data/messidor2/images /data/eyepacs/images \
  --csv_paths /data/messidor2/labels.csv /data/eyepacs/trainLabels.csv \
  --output_dir ./checkpoints/mdg_mix

Evaluation

python diabetic_retinopathy/evaluate.py \
  --model_path ./checkpoints/aptos \
  --test_dataset eyepacs \
  --test_data_path /path/to/eyepacs/images \
  --test_csv_path /path/to/eyepacs/labels.csv

Pretrained Models

LoRA adapters available on Dropbox:

https://www.dropbox.com/scl/fo/e51mect061togeqdwq2wg/ALCG9_9IPHLExKlpghC_6ks?rlkey=4xrh2df0y3xnv5yph1pbc91am&st=jhpgto0w&dl=0

Citation

@inproceedings{banerjee2026_humanknowledge,
  title={Human Knowledge Integrated Multi-modal Learning for Single Source Domain Generalization},
  author={Banerjee, Ayan and Thakur, Kuntal and Gupta, Sandeep},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year={2026},
  url={https://openaccess.thecvf.com/content/WACV2026/html/Banerjee_Human_Knowledge_Integrated_Multi-modal_Learning_for_Single_Source_Domain_Generalization_WACV_2026_paper.html}
}

License

MIT License