Access to GRCD requires PhysioNet credentialed access

GRCD was trained on MIMIC-CXR-JPG and Chest ImaGenome, which are released under the PhysioNet Credentialed Health Data License. Access is granted only to users who hold active PhysioNet credentialing and have signed the data use agreements for both datasets. Requests are reviewed manually.

GRCD: Grounded Region Change Detection for Multi-Finding Chest X-Ray Pairs

GRCD generates grounded temporal radiology reports from a prior and a current chest X-ray. Each finding in the output carries a bounding box and a change label (worsened, improved, or stable). Several findings can appear in one report.

Paper: arXiv:2607.02719
Code: https://github.com/UTSA-VIRLab/GRCD
Project page: https://grcd-cxr.github.io

Access conditions

The weights were trained on credentialed-access data. To use them you must:

hold active PhysioNet credentialing;
have signed the data use agreements for MIMIC-CXR-JPG v2.1.0 and Chest ImaGenome v1.0.0;
not redistribute the weights to anyone who does not meet the same requirements.

Requests are approved manually after we check the PhysioNet username.

Files

File	Contents
`adapter_model.bin`	LoRA adapter (r=128, α=256) for the language model
`adapter_config.json`	LoRA configuration
`non_lora_trainables.bin`	Vision-language projector, Region-Guided Change Token (RGCT) module, and gated cross-attention weights
`config.json`	Model configuration (dual pathway: top-6 prepended tokens plus cross-attention every 4th layer)
`SHA256SUMS`	Checksums for the weight files

The base LLM (lmsys/vicuna-7b-v1.5) and the BioViL-T image encoder are downloaded from their public sources at load time.

Usage

git clone https://github.com/UTSA-VIRLab/GRCD && cd GRCD
pip install -r requirements.txt && pip install -e .
huggingface-cli download aranya059/GRCD --local-dir ./checkpoints/grcd
python evaluation/eval_main.py --model_path ./checkpoints/grcd --test_path ./data/test.json

Regenerate the benchmark (data/test.json) yourself with the pipeline in the repository. The benchmark is not redistributed.

Results

7,967-pair multi-finding test set, mean over 3 seeds:

BLEU-4	METEOR	ROUGE-L	CheXbert F1	RadGraph F1	Report-level change detection
0.398	0.494	0.493	0.567	0.473	53.3%

Training

Data: 28,155 cleaned temporal pairs from MIMIC-CXR-JPG v2.1.0 and Chest ImaGenome v1.0.0, using the official patient-level splits.
Encoder: frozen BioViL-T.
Trained parts: LoRA on Vicuna-7B-v1.5, the projector, RGCT, and the cross-attention.
Schedule: two stages, 5 epochs each.

Intended use and limitations

For research on temporal report generation and grounding only. Not for clinical use.
Labels are silver-standard (derived by NLP). Change direction, especially improved vs. worsened, is the main error mode.
Models trained on patient data can memorize parts of their training data. Do not probe the model to recover report text or patient information.

Licence

These weights are released for research use under the conditions above. Use of the base models is also subject to their licences: Vicuna/Llama 2 and BioViL-T.

Citation

@article{aranya2026grcd,
  title   = {GRCD: Grounded Region Change Detection for Multi-Finding Chest X-Ray Pairs},
  author  = {Aranya, OFM Riaz Rahman and Najafirad, Peyman and Desai, Kevin},
  journal = {arXiv preprint arXiv:2607.02719},
  year    = {2026}
}

Downloads last month: -

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for aranya059/GRCD

Base model

lmsys/vicuna-7b-v1.5

Adapter

(186)

this model

Paper for aranya059/GRCD

GRCD: Grounded Region Change Detection for Multi-Finding Chest X-Ray Pairs

Paper • 2607.02719 • Published Jul 2