RadImageNet grants you permission, upon your agreeing to the terms of the Research Use Agreement, to view and use the Dataset for personal, non-commercial (e.g., academic) research purposes only. Any commercial use, sale, or other monetization, by you or your affiliates, is strictly prohibited under any and all circumstances.
Other than any limited rights expressly granted herein to you, RadImageNet retains all rights, title, and interest in the Dataset.
You may make a verbatim copy of the Dataset for non-commercial research use as permitted in the Research Use Agreement. You may not alter this verbatim copy for any reason. If another user within your organization wishes to use the Dataset, they must register as an individual user and comply with all the terms of the Research Use Agreement.
YOU MAY NOT DISTRIBUTE, PUBLISH, OR REPRODUCE A COPY of any portion, including the entirety, of the Dataset to anyone without express and specific prior written permission from RadImageNet.
YOU MAY NOT SHARE THE DOWNLOAD LINK to the Dataset with others. For example, if someone other than you within your organization wishes to use or view the Dataset, they must register as an individual user and agree to and comply with all the terms of the Research Use Agreement.
You must not modify, reverse engineer, decompile, or create derivative works from the Dataset. You must not remove or alter any copyright or other proprietary notices in the Dataset.
The Dataset has not been reviewed or approved by the Food and Drug Administration, or any other regulatory agency of the United States of America. The Dataset is being provided to you strictly and only for non-clinical, research use. In no event shall data or images generated through the use, directly or indirectly, in whole or in part, of the Dataset be used or relied upon in the diagnosis or provision of patient care. This Research Use Agreement expressly forbids the use, directly or indirectly, in whole or in part, of the Dataset in the diagnosis or provision of patient care.
THE DATASET IS PROVIDED “AS IS,” AND RADIMAGENET AND ITS COLLABORATORS MAKE NO WARRANTY, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO WARRANTIES OF MERCHANTABILITY AND FITNESS FOR ANY PARTICULAR PURPOSE,2 NOR DO THEY ASSUME ANY LIABILITY OR RESPONSIBILITY FOR THE USE OF THE DATASET.
You will not attempt to identify or re-identify any of the individual data subjects (e.g., patients). Identification or re-identification of individuals is strictly prohibited. Any identification or re-identification of any individual data subject shall be immediately reported to RadImageNet and may be subject to immediate termination of the use of the Dataset.
Any violation of the Research Use Agreement or other impermissible use shall be grounds for immediate termination of use of the Dataset. It is your duty to promptly report to RadImageNet any knowledge of any violation at any time. In the event that RadImageNet determines that you have violated this Research Use Agreement or made other impermissible use of the Dataset, RadImageNet may direct that you immediately return all copies of the Dataset and retain no copies thereof. RadImageNet may do this even if you did not cause the violation or impermissible use.

In consideration for your agreement to the terms and conditions contained in the Research Use Agreement, RadImageNet grants you limited permission to view and use the Dataset for personal, non-commercial research, as described herein. You may not otherwise copy, reproduce, retransmit, distribute, publish, commercially exploit or otherwise transfer any material from or related to the Dataset.

Limitation of Use

You may use the Dataset for legal purposes only.

Indemnification

You agree to indemnify and hold RadImageNet harmless from and not liable in any way for any claims, losses or damages, including legal fees, arising out of or resulting from your use of the Dataset or your violation or role in violation of the Research Use Agreement. You agree to fully cooperate in RadImageNet’s defense against any such claims. These terms and all other terms of the Research Use Agreement shall be governed by and interpreted in accordance with the laws of New York State.

RadImageNet-VQA: A Large-Scale CT and MRI Dataset for Radiologic Visual Question Answering

📖 Paper

Dataset Details

We introduce RadImageNet-VQA, a large-scale dataset designed for training and benchmarking radiologic VQA on CT and MRI exams. Built from the CT/MRI subset of RadImageNet and its expert-curated anatomical and pathological annotations, RadImageNet-VQA provides 750K images with 7.5M generated samples, including 750K medical captions for visual-text alignment and 6.75M question-answer pairs that span three radiology tasks: fine-grained pathology identification, anatomy recognition, and abnormality detection. The dataset includes open-ended, closed-ended, and multiple-choice questions across 8 anatomical regions and 97 pathologies, generated with prompt-based templates and constructed to probe visual-grounded understanding while minimizing text-only shortcut answering. For evaluation, we construct a stratified benchmark of 1,000 images with 9,000 question-answer pairs covering all tasks and question types.

Data Creation

RadImageNet-VQA was created to challenge multimodal models with tasks that demand radiology text-image understanding, pushing the boundaries of what these models can achieve in terms of perception and reasoning. The data for the RadImageNet-VQA dataset was build upon RadImageNet, a large expert-annotated medical imaging dataset in which each image is associated with a modality (CT, MRI, US), a body part (e.g., abdomen, hip, brain) and a pathology label. From this resource, we use the CT and MRI subsets to form the basis for generating clinically meaningful captions and VQA samples across anatomy, abnormality, and fine-grained pathology tasks.

Zero-shot Results

Zero-shot accuracies (%) of VLMs on RadImageNet-VQA benchmark. Results are reported across anatomy recognition, abnormality detection (Abn), and pathology identification using four question formats: Open (free-form), Closed+ (always 'yes' as true answer), Closed– (always 'no'), and MC (multiple-choice).

Model		Anatomy			Abnormality		Pathology			Average
	Open	Closed+	Closed–	MC	Closed	Open	Closed+	Closed–	MC
General-purpose models
LLaVA-OneVision-Qwen2-7B	48.4	82.7	81.3	88.7	49.8	16.0	55.3	61.3	33.6	57.5
Qwen2.5-VL-3B-Instruct	37.7	83.7	77.1	77.9	70.5	10.0	78.1	21.4	34.8	54.6
Qwen2.5-VL-7B-Instruct	37.5	84.9	79.1	80.5	69.5	9.8	69.2	47.4	30.1	56.4
InternVL3.5-8B	50.9	98.1	75.9	93.3	58.9	9.9	85.9	27.8	41.8	60.3
InternVL3.5-14B	56.6	98.2	74.4	89.9	74.4	11.7	86.7	33.7	47.1	63.6
GPT-5	44.3	72.4	81.8	89.3	27.5	15.8	54.9	68.3	41.2	54.9
Gemini 2.5 Pro	65.7	76.5	81.9	88.8	17.8	21.1	50.2	30.1	44.4	52.9
Medical-specialized models
LLaVA-Med-v1.5-mistral-7b	44.3	89.9	55.3	58.1	22.4	10.2	41.8	66.6	26.4	48.2
HuatuoGPT-Vision-7B	45.4	82.5	89.0	88.3	60.6	13.6	65.5	69.2	44.6	48.9
medgemma-4b-it	62.9	76.4	82.5	84.8	55.4	30.6	54.2	77.4	36.8	51.5
Lingshu-7B	49.6	90.7	85.1	88.9	47.9	15.7	57.0	78.8	29.6	60.4
Lingshu-32B	45.2	75.5	92.1	89.3	54.5	14.4	46.4	88.8	31.7	59.8

Bold = best, italic = second best

Data Structure

Alignment Data

The alignment component contains single caption samples per image, intended to align visual content with concise clinical descriptions.

Each instance conceptually includes:

an image
a single prompt–response pair
structured metadata

Fields:

id: unique sample identifier
image: relative path to the medical image
conversations: one human prompt and one descriptive response
metadata: modality, anatomical location, abnormality flag, pathology label

The response provides a brief clinical description of the image.

Instruction Data

The instruction component contains multiple question–answer pairs per image and is intended for instruction tuning of multimodal models.

Each instance includes:

an image
one or more QA-style conversation turns
structured metadata describing the task

Supported instruction types include image description, pathology identification, modality recognition, and anatomical localization.

Benchmark Data

The benchmark split is designed for standardized evaluation of medical VQA models.

It contains 9,000 question–answer pairs across 1,000 images and includes three question types:

open-ended (free-form answers)
closed-ended (yes/no)
multiple-choice (options A–D)

Benchmark fields:

image: medical image reference
question: question presented to the model
choices: answer options (multiple-choice only)
answer: ground-truth answer
question_type: open, yes/no, or multiple-choice
metadata: modality, anatomy, pathology, and correctness labels

Metadata

Metadata fields provide structured clinical and contextual information:

modality: imaging modality (e.g., CT, MRI)
location: anatomical region
is_abnormal: presence of pathology
pathology: pathology category
content_type: task type (description, pathology, etc.)
question_id: question template identifier
correct_text: textual form of the correct answer (when applicable)

Data Splits

The dataset is organized into three configurations with training and validation splits:

	Alignment		Instruction Tuning		Benchmark
	Train	Validation	Train	Validation	Test
Samples	750,009	83,668	750,009	83,668	9,000
Images	750,009	83,668	750,009	83,668	1,000
QAs per image	1	1	~9	~9	9
Total QAs	750K	83K	6.75M	753K	9K

Acknowledgments

The dataset is built upon RadImageNet https://www.radimagenet.com/.

Citation

@inproceedings{
butsanets2025radimagenetvqa,
title={RadImageNet{VQA}: A Large-Scale {CT} and {MRI} Dataset for Medical Visual Question Answering},
author={L{\'e}o Butsanets and Charles Corbi{\`e}re and Julien Khlaut and Pierre Manceron and Corentin Dancette},
year={2025},
url={https://openreview.net/forum?id=khHKvZ9sLD},
}

Downloads last month: 59