Saved in:
| Main Authors: | , , , , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2605.20525 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866917514562240512 |
|---|---|
| author | Abbasi, Mohammad H. Nerrise, Favour Ghosh, Shaurnav Yesiloglu, Ridvan Mao, Yuncong Trang, Bailey Asadi, Mohammad Daniel, Merryn Kung, Gustavo Chau Loo Chang, Ken Shah, Pavan Pinkesh Turnbull, Adam Younes, Kyan Dehkharghani, Seena Adeli, Ehsan |
| author_facet | Abbasi, Mohammad H. Nerrise, Favour Ghosh, Shaurnav Yesiloglu, Ridvan Mao, Yuncong Trang, Bailey Asadi, Mohammad Daniel, Merryn Kung, Gustavo Chau Loo Chang, Ken Shah, Pavan Pinkesh Turnbull, Adam Younes, Kyan Dehkharghani, Seena Adeli, Ehsan |
| contents | We present NeuroQA, a large-scale benchmark for visual question answering in 3D brain magnetic resonance imaging (MRI), with 56,953 QA pairs from 12,977 subjects across 12 datasets. It spans ages 5-104 and five clinical domains: Alzheimer's, Parkinson's, tumors, white matter disease, and neurodevelopment. Unlike prior medical Visual Question Answering (VQA) efforts that operate on 2D slices or rely on narrow diagnostic labels, NeuroQA pairs every item with a full 3D volume. It evaluates 11 clinically grounded reasoning skills across Yes/No, multiple-choice, and open-ended formats. Of the 203 templates, 131 are image-grounded (answerable from a 3-plane viewer) and 72 are image-informed (ground truth from quantitative volumetry or clinical instruments). To remove text-only shortcuts, we apply answer-distribution refinement, reducing closed-format text-only accuracy from $>$80% to 44.6%; image necessity is assessed separately through an image-grounding protocol released with the benchmark. A 38-rule deterministic pipeline and two rounds of expert review verify every QA pair against FreeSurfer measurements, metadata, or radiology report fields, with zero same-subject contradictions across templates. We conduct a clinician evaluation in which two clinicians independently assess 100 frozen test items on a three-plane viewer. On closed-format (Yes/No + multiple-choice) test-public items, the best zero-shot vision-language model and a supervised 3D CNN baseline reach 47.5% and 43.7% accuracy respectively, both below the 49.4% text-only majority-template floor. NeuroQA adopts a two-tier release with public QA pairs for open-access datasets and reproducible generation scripts for datasets restricted by data use agreements (DUAs), plus subject-level splits, a held-out private test set, and an online leaderboard. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_20525 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | NeuroQA: A Large-Scale Image-Grounded Benchmark for 3D Brain MRI Understanding Abbasi, Mohammad H. Nerrise, Favour Ghosh, Shaurnav Yesiloglu, Ridvan Mao, Yuncong Trang, Bailey Asadi, Mohammad Daniel, Merryn Kung, Gustavo Chau Loo Chang, Ken Shah, Pavan Pinkesh Turnbull, Adam Younes, Kyan Dehkharghani, Seena Adeli, Ehsan Computer Vision and Pattern Recognition Artificial Intelligence Computation and Language Machine Learning Image and Video Processing We present NeuroQA, a large-scale benchmark for visual question answering in 3D brain magnetic resonance imaging (MRI), with 56,953 QA pairs from 12,977 subjects across 12 datasets. It spans ages 5-104 and five clinical domains: Alzheimer's, Parkinson's, tumors, white matter disease, and neurodevelopment. Unlike prior medical Visual Question Answering (VQA) efforts that operate on 2D slices or rely on narrow diagnostic labels, NeuroQA pairs every item with a full 3D volume. It evaluates 11 clinically grounded reasoning skills across Yes/No, multiple-choice, and open-ended formats. Of the 203 templates, 131 are image-grounded (answerable from a 3-plane viewer) and 72 are image-informed (ground truth from quantitative volumetry or clinical instruments). To remove text-only shortcuts, we apply answer-distribution refinement, reducing closed-format text-only accuracy from $>$80% to 44.6%; image necessity is assessed separately through an image-grounding protocol released with the benchmark. A 38-rule deterministic pipeline and two rounds of expert review verify every QA pair against FreeSurfer measurements, metadata, or radiology report fields, with zero same-subject contradictions across templates. We conduct a clinician evaluation in which two clinicians independently assess 100 frozen test items on a three-plane viewer. On closed-format (Yes/No + multiple-choice) test-public items, the best zero-shot vision-language model and a supervised 3D CNN baseline reach 47.5% and 43.7% accuracy respectively, both below the 49.4% text-only majority-template floor. NeuroQA adopts a two-tier release with public QA pairs for open-access datasets and reproducible generation scripts for datasets restricted by data use agreements (DUAs), plus subject-level splits, a held-out private test set, and an online leaderboard. |
| title | NeuroQA: A Large-Scale Image-Grounded Benchmark for 3D Brain MRI Understanding |
| topic | Computer Vision and Pattern Recognition Artificial Intelligence Computation and Language Machine Learning Image and Video Processing |
| url | https://arxiv.org/abs/2605.20525 |