NeuroQA: A Large-Scale Image-Grounded Benchmark for 3D Brain MRI Understanding

NeuroQA is a benchmark for visual question answering on 3D brain MRI volumes. The full benchmark contains 56,953 QA pairs over 12,977 subjects from 12 neuroimaging datasets, evaluating 11 clinically grounded reasoning skills in three answer formats (Yes/No, multiple-choice, open-ended).

This HuggingFace repository hosts the Tier 1 (Public) subset of NeuroQA: 19,262 QA pairs drawn from six permissively licensed datasets. The remaining Tier 2 (DUA) subset must be reproduced by users from their own data access — see the GitHub repository for the deterministic generation pipeline.

Quick links

Resource	Where
Code, scripts, evaluation toolkit	github.com/mhabbasiit/neuroqa
Tier 2 (DUA) generation pipeline	github.com/mhabbasiit/neuroqa — tier2_pipeline/
Image preprocessing pipeline (skullstrip + MNI registration)	github.com/mhabbasiit/sMRI
Leaderboard & test-private submissions	neuroqa.stanford.edu

---

What's in this repository

mhabbasiit/NeuroQA/
├── all_public/              # combined splits (6 public datasets)
│   ├── train.json           # 13,598 items
│   ├── val.json             #  1,892 items
│   └── test.json            #  1,813 items   (= test_public, public-only subset)
└── per_dataset/             # per-dataset breakdown
    ├── BraTS-GLI/{train,val,test}.json   # 2,787 items
    ├── BraTS-MEN/{train,val,test}.json   # 2,418 items
    ├── CC359/{train,val,test}.json       # 3,092 items
    ├── HCP-YA/{train,val,test}.json      # 5,909 items
    ├── IXI/{train,val,test}.json         # 2,933 items
    └── WMH/{train,val,test}.json         #   164 items

Total Tier 1 on HuggingFace: 17,303 QA pairs (Train + Val + Test-Public). The remaining 1,959 Tier 1 items belong to the held-out test-private split, which is not released here; they are kept for the public leaderboard.

---

Two-tier release

NeuroQA's source datasets have non-uniform licensing, so the benchmark is released in two tiers:

	Tier 1 (Public)	Tier 2 (DUA-restricted)
Datasets	BraTS-GLI, BraTS-MEN, CC359, HCP-YA, IXI, WMH	ABCD, ADNI, AIBL, HCP-Aging, HCP-Development, PPMI
QA pairs	19,262	37,691
Subjects	2,467	10,510
Distribution	This HuggingFace dataset (QA only)	Generation scripts on GitHub — users obtain raw data themselves and reproduce QA bit-for-bit

Both tiers share identical templates, validation rules, and evaluation code, so any claim established on Tier 1 transfers to Tier 2.

---

Image data — not included

This repository contains only QA JSON files. All image_path, image_path_t2, and image_path_prior fields are set to "#".

To use NeuroQA you need to:

1. Download raw images from each public source:

   • BraTS-GLI / BraTS-MEN: synapse.org/brats
   • CC359: sites.google.com/view/calgary-campinas-dataset
   • HCP-YA: humanconnectome.org
   • IXI: brain-development.org/ixi-dataset
   • WMH: wmh.isi.uu.nl

2. Preprocess to MNI-152 space using our deterministic pipeline: 👉 github.com/mhabbasiit/sMRI This applies skullstrip (FreeSurfer SynthStrip), rigid + affine registration to MNI-152, and z-score normalization with cropping. Running this pipeline ensures voxel-for-voxel consistency with the images NeuroQA was built on.

3. Map the placeholder "#" paths in the QA JSONs to your local preprocessed images by subject_id + session_date.

---

Loading the dataset

from datasets import load_dataset

# Combined public (6 datasets)
ds = load_dataset("mhabbasiit/NeuroQA", data_files={
    "train": "all_public/train.json",
    "val":   "all_public/val.json",
    "test":  "all_public/test.json",
})

# Single dataset
ds_brats = load_dataset("mhabbasiit/NeuroQA", data_files={
    "train": "per_dataset/BraTS-GLI/train.json",
    "val":   "per_dataset/BraTS-GLI/val.json",
    "test":  "per_dataset/BraTS-GLI/test.json",
})

Or load directly with json:

import json
items = json.load(open("all_public/test.json"))
# Each item has: neuroqa_id, dataset, subject_id, question, answer, options,
#                question_type (YN/MCQ/Open), category, image_path, image_path_t2, ...

---

Item schema

Every QA item is a JSON object with these fields:

Field	Type	Description
neuroqa_id	str	Unique identifier (e.g., BraTS-GLI-00427)
dataset	str	Source dataset
subject_id	str	Subject identifier within the source dataset
session_date	str	Session/visit identifier
question	str	Question text
answer	str	Ground-truth answer
options	list/null	Four options (MCQ only); null for YN and Open
question_type	str	One of YN, MCQ, Open
category	str	One of 11 reasoning categories (Anatomy, Diagnosis, …)
image_dependency	str	image-grounded or image-informed
image_required	bool	Whether the image is needed to answer
image_path	str	Path to T1 volume (set to "#" in this release)
image_path_t2	str/null	Path to T2 volume if available (set to "#" if present)
image_path_prior	str/null	Path to prior scan (longitudinal items only)
answerable_modalities	list	Modalities for which the question is answerable
human_assessable	bool	Whether the answer is recoverable by visual inspection
tier	str	pub or dua
split	str	train, val, test_public, or test_private

---

Evaluation

Evaluation is performed with the official toolkit on GitHub:

git clone https://github.com/mhabbasiit/neuroqa
cd neuroqa
python eval/evaluate.py --predictions preds.json --ground-truth all_public/test.json

This computes per-dataset, per-category, per-format accuracy plus the Shortcut Score, BERTScore, and ROUGE-L for open-ended items. See eval/README.md on GitHub for details.

Test-private and the leaderboard

The held-out test-private split (5,581 items, including 1,959 from public datasets) has hidden ground-truth labels and is not released. To evaluate on test-private, submit predictions to the public leaderboard:

👉 neuroqa.stanford.edu

The leaderboard reports per-dataset, per-category, and per-format accuracy under standardized scoring.

---

License

QA pairs released here are licensed under CC-BY-4.0.

Note: the underlying source images are governed by the original dataset licenses (BraTS, IXI, WMH, CC359, HCP-YA), which users must obtain and respect separately. This repository does not distribute images.

---

Citation

If you use NeuroQA, please cite:

@misc{neuroqa2026,
  title        = {NeuroQA: A Large-Scale Image-Grounded Benchmark for 3D Brain MRI Understanding},
  author       = {Abbasi, Mohammad H. and Nerrise, Favour and others},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/mhabbasiit/NeuroQA}},
  note         = {Mohammad H. Abbasi and Favour Nerrise contributed equally.}
}

---

Contact

Questions about the dataset: open an issue on github.com/mhabbasiit/neuroqa.

Leaderboard / test-private requests: neuroqa.stanford.edu.