iamlkx/vqa-rad

---
license: cc0-1.0
task_categories:
- visual-question-answering
language:
- en
paperswithcode_id: vqa-rad
tags:
- medical
pretty_name: VQA-RAD
size_categories:
- 1K<n<10K
dataset_info:
features:
- name: image
dtype: image
- name: question
dtype: string
- name: answer
dtype: string
splits:
- name: train
num_bytes: 95883938.139
num_examples: 1793
- name: test
num_bytes: 23818877.0
num_examples: 451
download_size: 34496718
dataset_size: 119702815.139
---

# Dataset Card for VQA-RAD

## Dataset Description
VQA-RAD is a dataset of question-answer pairs on radiology images. The dataset is intended to be used for training and testing
Medical Visual Question Answering (VQA) systems. The dataset includes both open-ended questions and binary "yes/no" questions.
The dataset is built from [MedPix](https://medpix.nlm.nih.gov/), which is a free open-access online database of medical images.
The question-answer pairs were manually generated by a team of clinicians.

**Homepage:** [Open Science Framework Homepage](https://osf.io/89kps/)<br>
**Paper:** [A dataset of clinically generated visual questions and answers about radiology images](https://www.nature.com/articles/sdata2018251)<br>
**Leaderboard:** [Papers with Code Leaderboard](https://paperswithcode.com/sota/medical-visual-question-answering-on-vqa-rad)

### Dataset Summary
The dataset was downloaded from the [Open Science Framework Homepage](https://osf.io/89kps/) on June 3, 2023. The dataset contains
2,248 question-answer pairs and 315 images. Out of the 315 images, 314 images are referenced by a question-answer pair, while 1 image
is not used. The training set contains 3 duplicate image-question-answer triplets. The training set also has 1 image-question-answer
triplet in common with the test set. After dropping these 4 image-question-answer triplets from the training set, the dataset contains
2,244 question-answer pairs on 314 images.

#### Supported Tasks and Leaderboards
This dataset has an active leaderboard on [Papers with Code](https://paperswithcode.com/sota/medical-visual-question-answering-on-vqa-rad)
where models are ranked based on three metrics: "Close-ended Accuracy", "Open-ended accuracy" and "Overall accuracy". "Close-ended Accuracy" is
the accuracy of a model's generated answers for the subset of binary "yes/no" questions. "Open-ended accuracy" is the accuracy
of a model's generated answers for the subset of open-ended questions. "Overall accuracy" is the accuracy of a model's generated
answers across all questions.

#### Languages
The question-answer pairs are in English.

## Dataset Structure

### Data Instances
Each instance consists of an image-question-answer triplet.
```
{
'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=566x555>,
'question': 'are regions of the brain infarcted?',
'answer': 'yes'
}
```
### Data Fields
- `'image'`: the image referenced by the question-answer pair.
- `'question'`: the question about the image.
- `'answer'`: the expected answer.

### Data Splits
The dataset is split into training and test. The split is provided directly by the authors.

| | Training Set | Test Set |
|-------------------------|:------------:|:---------:|
| QAs |1,793 |451 |
| Images |313 |203 |

## Additional Information

### Licensing Information
The authors have released the dataset under the CC0 1.0 Universal License.

### Citation Information
```
@article{lau2018dataset,
title={A dataset of clinically generated visual questions and answers about radiology images},
author={Lau, Jason J and Gayen, Soumya and Ben Abacha, Asma and Demner-Fushman, Dina},
journal={Scientific data},
volume={5},
number={1},
pages={1--10},
year={2018},
publisher={Nature Publishing Group}
}
```