README.md

---
license: cc-by-4.0
task_categories:
- image-to-text
language:
- en
tags:
- image
- text
- medical-imaging
- radiology
- chest-x-ray
- radiology-report-generation
- reinforcement-learning
- iu-xray
- mimic-cxr
- arxiv:2607.10147
pretty_name: REVA-PO Datasets
---

# REVA-PO Datasets

Dataset files and annotations used in **REVA-PO: Stabilizing Reinforcement Learning for Chest X-ray Report Generation**, accepted to ECCV 2026.

[![Paper](https://img.shields.io/badge/arXiv-2607.10147-b31b1b.svg)](https://arxiv.org/abs/2607.10147)
[![Code](https://img.shields.io/badge/GitHub-REVA--PO-181717.svg?logo=github)](https://github.com/LiGuo12/REVA_PO)
[![Weights](https://img.shields.io/badge/Weights-Hugging_Face-FFD21E.svg)](https://huggingface.co/liguo12/REVA_PO_Weights)

## Overview

REVA-PO is a reinforcement learning framework for chest X-ray report generation. It uses Response-Weighted Regularization and Validation-Anchored Policy Reset to improve training stability and exploration.

This repository contains the IU-Xray data package and the MIMIC-CXR annotation files used by the released code. The model checkpoints are available in the [REVA-PO weights repository](https://huggingface.co/liguo12/REVA_PO_Weights).

## Repository Contents

```text
REVA_PO_Datasets/
├── iuxray_dataset/
│   ├── annotation_with_categories.json
│   └── images.zip
└── mimic_dataset/
    ├── mimic_with_categories.json
    └── mimic_with_categories_sampled_10k.json
```

| Path | Description |
|---|---|
| `iuxray_dataset/images.zip` | IU-Xray chest X-ray images arranged for the released REVA-PO data loader. |
| `iuxray_dataset/annotation_with_categories.json` | IU-Xray train, validation, and test annotations with report text and clinical categories. |
| `mimic_dataset/mimic_with_categories.json` | Full MIMIC-CXR annotation file used for supervised training. |
| `mimic_dataset/mimic_with_categories_sampled_10k.json` | MIMIC-CXR annotation file with 9,974 randomly sampled training instances for reinforcement learning. The validation and test splits are unchanged. |

## Download

Install the Hugging Face Hub client:

```bash
pip install -U huggingface_hub
```

Download the full repository:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="liguo12/REVA_PO_Datasets",
    repo_type="dataset",
    local_dir="REVA_PO_Datasets",
)
```

To download only selected files:

```python
from huggingface_hub import hf_hub_download

iu_annotations = hf_hub_download(
    repo_id="liguo12/REVA_PO_Datasets",
    repo_type="dataset",
    filename="iuxray_dataset/annotation_with_categories.json",
)

mimic_annotations = hf_hub_download(
    repo_id="liguo12/REVA_PO_Datasets",
    repo_type="dataset",
    filename="mimic_dataset/mimic_with_categories_sampled_10k.json",
)
```

## IU-Xray Setup

Extract `images.zip` inside `iuxray_dataset/`:

```bash
cd REVA_PO_Datasets/iuxray_dataset
unzip images.zip
```

The resulting directory should be:

```text
iuxray_dataset/
├── images/
│   ├── CXR2384_IM-0942/
│   │   ├── 0.png
│   │   └── 1.png
│   └── ...
└── annotation_with_categories.json
```

Set the following fields in the REVA-PO configuration files:

```yaml
storage: /path/to/iuxray_dataset
ann_file: /path/to/iuxray_dataset/annotation_with_categories.json
```

## MIMIC-CXR Setup

MIMIC-CXR images are not included in this repository. Obtain authorized access to [MIMIC-CXR-JPG v2.1.0](https://physionet.org/content/mimic-cxr-jpg/2.1.0/) through PhysioNet, then place its `files/` directory beside the downloaded annotation files:

```text
mimic_dataset/
├── files/
│   ├── p10/
│   │   ├── p10000032/
│   │   │   ├── s50414267/
│   │   │   │   ├── 02aa804e-bde0afdd-112c0b34-7bc16630-4e384014.jpg
│   │   │   │   └── ...
│   │   │   └── ...
│   │   └── ...
│   └── ...
├── mimic_with_categories.json
└── mimic_with_categories_sampled_10k.json
```

For supervised training (stage 1 and stage 2), use:

```yaml
storage: /path/to/mimic_dataset
ann_file: /path/to/mimic_dataset/mimic_with_categories.json
```

For the released reinforcement learning (stage 3), use:

```yaml
storage: /path/to/mimic_dataset
ann_file: /path/to/mimic_dataset/mimic_with_categories_sampled_10k.json
```

## IU-Xray Annotation Format

The annotation files are JSON objects organized by data split:

```json
{
  "train": [
    {
      "id": "CXR2384_IM-0942",
      "report": "The heart size and pulmonary vascularity appear within normal limits...",
      "image_path": [
        "CXR2384_IM-0942/0.png",
        "CXR2384_IM-0942/1.png"
      ],
      "split": "train",
      "positive_categories": ["No Finding"],
      "uncertain_categories": [],
      "negative_categories": [
        "Cardiomegaly",
        "Lung Opacity",
        "Pneumothorax",
        "Pleural Effusion"
      ],
      "categories": [
        "No Finding"
      ]
    }
  ],
  "val": [...],
  "test": [...]
}
```

Common fields include:

| Field | Description |
|---|---|
| `id` | Study or examination identifier. |
| `report` | Reference radiology report. |
| `image_path` | Relative path or paths to the associated chest X-ray images. |
| `split` | Data split. |
| `positive_categories` | Clinical categories labeled as positive. |
| `uncertain_categories` | Clinical categories labeled as uncertain. |
| `negative_categories` | Clinical categories labeled as negative. |
| `categories` | Categories used by the REVA-PO training pipeline. |

The exact fields can differ slightly between IU-Xray and MIMIC-CXR. Use the released data loaders in the [code repository](https://github.com/LiGuo12/REVA_PO) as the reference implementation.

## Intended Use

These files are intended for research on:

- Chest X-ray report generation
- Medical image-to-text generation
- Clinical vision-language learning
- Reinforcement learning for report generation
- Evaluation of linguistic quality and clinical accuracy

They are not intended for direct clinical diagnosis, treatment decisions, or use without independent medical validation.

## Limitations

The annotations inherit the limitations of IU-Xray and MIMIC-CXR, including dataset-specific reporting patterns, label noise, class imbalance, and differences between institutions. Performance measured on these datasets does not establish safety or effectiveness in clinical use.

## License and Access Notes

The annotation files and other original materials released in this repository are licensed under the [Creative Commons Attribution 4.0 International License](https://creativecommons.org/licenses/by/4.0/). You may use, share, and adapt these materials for research or other purposes, provided that appropriate credit is given to the REVA-PO authors.

This license applies only to materials created and released by the REVA-PO authors. It does not replace the terms of the original datasets:

- MIMIC-CXR-JPG requires credentialed access through PhysioNet. Users must follow the PhysioNet Credentialed Health Data Use Agreement and the MIMIC-CXR-JPG access conditions.
- IU-Xray data remains subject to the terms of its original source.
- The REVA-PO source code is released under the license stated in the [GitHub repository](https://github.com/LiGuo12/REVA_PO).
- Users are responsible for checking and following all terms that apply to the source datasets.

## Citation

If you use these files or the REVA-PO method, please cite:

```bibtex
@misc{guo2026revapo,
  title        = {REVA-PO: Stabilizing Reinforcement Learning for Chest X-ray Report Generation},
  author       = {Li Guo and Anas M. Tahir and Z. Jane Wang},
  year         = {2026},
  eprint       = {2607.10147},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url          = {https://arxiv.org/abs/2607.10147}
}
```

Please also cite the original IU-Xray and MIMIC-CXR resources when applicable.