Linn5412/CXR-GroundVQA

---
license: mit
task_categories:
- visual-question-answering
language:
- en
tags:
- chest-x-ray
- phrase-grounding
- medical-imaging
- synthetic-data
- VQA
size_categories:
- 1K<n<10K
---

# CXR-GroundVQA

A synthetic VQA dataset for chest X-ray phrase grounding, containing 6,006 question-answer pairs across 995 radiologist-annotated CXR images from VinDr-CXR.

## Overview

| Property | Value |
|----------|-------|
| Total QA pairs | 6,006 |
| Images | 995 (from VinDr-CXR training split) |
| Question types | 8 |
| Finding categories | 22 |
| Bbox format | Normalized [x, y, w, h] in [0, 1] |
| Language | English |
| Quality score | Mean 0.986, 100% verification pass |

## Question Types

| Category | Type | Count | % |
|----------|------|-------|---|
| Bbox-localization | Open-ended Localization | 836 | 13.9 |
| Bbox-localization | Single-choice Localization | 673 | 11.2 |
| Bbox-localization | Multi-choice Localization | 446 | 7.4 |
| Bbox-localization | Zero-knowledge Detection | 481 | 8.0 |
| Label-identification | Open-ended Identification | 1,092 | 18.2 |
| Label-identification | Single-choice Identification | 836 | 13.9 |
| Label-identification | Multi-choice Identification | 419 | 7.0 |
| Polarity | Polarity Judgment | 1,223 | 20.4 |

## Data Format

Each sample is a JSON object with messages (conversation turns) and images (file paths):

```json
{
"messages": [
{
"role": "user",
"content": "<image>
What abnormality can be observed within the region designated by [0.735, 0.565, 0.035, 0.038]?"
},
{
"role": "assistant",
"content": "The region [0.735, 0.565, 0.035, 0.038] shows Nodule/Mass."
}
],
"images": [
"images/0021df30f3fddef551eb3df4354b1d06.png"
]
}
```

Bounding boxes use normalized [x, y, w, h] coordinates where (x, y) is the top-left corner, and (w, h) are width and height relative to image dimensions, all in [0, 1].

## Image Setup

**Images are not included in this repository** due to the PhysioNet Credentialed Health Data License of VinDr-CXR. To use this dataset:

1. Obtain credentialed access to [VinDr-CXR v1.0.0](https://physionet.org/content/vindr-cxr/1.0.0/) on PhysioNet.
2. Download the training split DICOM files.
3. Convert the 995 referenced DICOM files to PNG format and place them in an images/ directory.

The image field in each sample specifies the expected relative path (e.g., images/xxxx.png).

## Source

Generated from the training split of [VinDr-CXR](https://physionet.org/content/vindr-cxr/1.0.0/) using an automated LLM-driven pipeline with placeholder-based factual grounding and 7-check quality assurance. See the paper for details.

## Citation

```bibtex
@inproceedings{lin2026cxrgroundvqa,
title={CXR-GroundVQA: An LLM-Synthesized VQA Dataset for Chest X-ray Phrase Grounding},
author={Lin, Jiaming},
booktitle={ACM Multimedia},
year={2026}
}
```