NeuroQA: Large-Scale 3D Brain MRI Benchmark for Visual Question Answering
NeuroQA has been launched by researchers as a comprehensive benchmark for visual question answering using 3D brain MRI, featuring 56,953 question-answer pairs sourced from 12,977 individuals across 12 datasets. This dataset encompasses ages ranging from 5 to 104 and includes five clinical areas: Alzheimer's, Parkinson's, tumors, white matter disease, and neurodevelopment. In contrast to previous medical VQA initiatives that utilized 2D slices or limited diagnostic labels, NeuroQA associates each question with a complete 3D volume. It tests 11 clinically relevant reasoning abilities through Yes/No, multiple-choice, and open-ended questions. Out of 203 templates, 131 are image-grounded, while 72 are image-informed. To eliminate reliance on text-only shortcuts, the accuracy of closed-format text-only responses was decreased from over 80% to 44.6%, with the necessity of images evaluated independently.
Key facts
- NeuroQA includes 56,953 QA pairs from 12,977 subjects across 12 datasets.
- Subjects range in age from 5 to 104 years.
- Covers five clinical domains: Alzheimer's, Parkinson's, tumors, white matter disease, and neurodevelopment.
- Each item is paired with a full 3D volume, unlike prior 2D-slice approaches.
- Evaluates 11 clinically grounded reasoning skills.
- Formats include Yes/No, multiple-choice, and open-ended.
- 203 templates total: 131 image-grounded, 72 image-informed.
- Answer-distribution refinement reduces text-only accuracy from >80% to 44.6%.
Entities
—