mlinslab/neurobench

Evaluating vision-language models on clinical neuroimaging

Overview

NeuroBench is a benchmark for evaluating vision-language models in neuroradiology. It uses full imaging studies comprising one or more 3D scans, with tasks designed to reflect clinical radiological interpretation. It includes study-level classification, closed-ended visual question answering, report generation, and red-team evaluations with misleading prompts.

This repository hosts NeuroBench’s images, question CSVs, and baseline reports generated by Neuro-VL, a multimodal neuroimaging model released alongside NeuroBench. The GitHub repository provides the evaluation code and scoring protocols.

Directory	Contents
data/	Image files in both NIfTI and preprocessed tensor formats
neuro_vl_assets/	Neuro-VL-generated baseline reports
zeroshot_classification/	CSVs for study-level diagnostic classification
visual_question_answering/closed_ended/	CSVs for multiple-choice and yes/no questions
visual_question_answering/report_generation/	CSVs for report generation
visual_question_answering/red_team/	CSVs for red-team evaluations

Download

You can download files directly from Hugging Face or use the NeuroBench CLI downloader.

Install the Hugging Face client:

python -m pip install huggingface_hub

Download raw images in NIfTI format (img.nii.gz or img.nii):

from pathlib import Path
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="mlinslab/neurobench",
    repo_type="dataset",
    allow_patterns="data/nifti/**",
    local_dir="./neurobench_data",
)

root = Path("./neurobench_data/data/nifti/rsna")
for batch in root.glob("batch_[0-4]"):
    for patient in batch.iterdir():
        patient.rename(root / patient.name)
    batch.rmdir()

Alternatively, you can download preprocessed image tensors, the format expected by the Neuro-VL model. See the demo script in the Neuro-VL repository for preprocessing details.

from pathlib import Path
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="mlinslab/neurobench",
    repo_type="dataset",
    allow_patterns="data/tensor/**",
    local_dir="./neurobench_data",
)

root = Path("./neurobench_data/data/tensor/rsna")
for batch in root.glob("batch_[0-4]"):
    for patient in batch.iterdir():
        patient.rename(root / patient.name)
    batch.rmdir()

Scan Paths and Formats

All question CSVs contain several common columns. The primary identifier columns are dataset, patient, and study. The files column lists scan directories based on the original source names for the images. types specifies a normalized description of each scan in the same order. Decode both with ast.literal_eval and preserve their order.

Choose the format root (data/tensor or data/nifti), then append the row's dataset value unchanged. The remaining path depends on the source collection:

Layout	Tensor path relative to the download root
Study directories, including BraTS and MR-RATE	data/tensor/<dataset>/<study>/<scan>/img.pt
Patient/study directories, including CFB-GBM, ISLES, and RSNA	data/tensor/<dataset>/<patient>/<study>/<scan>/img.pt

Here, <scan> is an entry in files.

NIfTIs are under the separate data/nifti/ root and use img.nii.gz or img.nii, where available.

Citation

NeuroBench was released alongside Neuro-VL: A multimodal intelligence for neuroimaging.

The Neuro-VL/NeuroBench preprint is coming soon.