VQA-RAD: Medical Imaging Questions and Answers
A medical VQA data set containing questions and answers generated by clinicians on images from radiology, intended to train AI models.
Description
The dataset VQA-RAD is a corpus of radiological images associated with 2,244 pairs of questions and answers generated manually by clinicians. It combines open-ended questions and binary “yes/no” questions, providing a solid foundation for training and evaluating Visual Question Answering (VQA) AI models in the medical field. The images come from the free MedPix database.
What is this dataset for?
- Training specialized VQA models in medicine
- Testing the understanding of medical images by AIs
- Contribute to AI-assisted diagnostic research
Can it be enriched or improved?
Yes, this dataset can be enriched with other imaging modalities (scanner, MRI), or extended with automatically generated questions. A possible improvement would be the categorization of questions by level of difficulty or medical specialty. It is also possible to associate clinical justifications or reports to reinforce contextual learning.
🔎 In summary
🧠 Recommended for
- Medical AI researchers
- VQA teams
- Radiology assistant projects
🔧 Compatible tools
- PyTorch
- Hugging Face Transformers
- MONAI
- Detectron2
💡 Tip
Use a strict train/test separation due to the presence of reported duplicates in the dataset.
Frequently Asked Questions
Is this dataset annotated by health professionals?
Yes, all questions and answers were generated manually by a team of clinicians, ensuring high semantic quality.
Is it suitable for training general question and answer models?
No, this dataset is specialized for medical imaging. For general use, use datasets like VQA v2 or GQA.
Can this dataset be combined with other medical sources?
Yes, it can be merged with other imagery datasets to create a larger or multilingual corpus, with attention to formats.
