mlinslab/neurobench

NeuroBench provides MRI/CT studies and evaluation tasks for benchmarking vision-language models on brain imaging interpretation.

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 preprocessed PyTorch tensors compatible with Neuro-VL:

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()

Alternatively, you can 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()

For Neuro-VL preprocessing details, see this script. See Neuro-VL for model inference and NeuroBench for evaluation tools, including LLM judges for report generation and red-teaming.

If you use NeuroBench, please consider citing our paper: A frontier machine intelligence for human neuroimaging.

@article{zhao2026frontier,
  title={A frontier machine intelligence for human neuroimaging},
  author={Chenhui Zhao and Akshay Rao and Akhil Kondepudi and Yiwei Lyu and Xinhai Hou and Rikhil Seshadri and Cheng Jiang and Soumyanil Banerjee and Rushikesh Joshi and Samir Harake and Ashok Srinivasan and Shruti Mishra and Francisco Rivas-Rodriguez and Seunghong Rhee and Nathaniel Roddenberry and Neeraj Chaudhary and Zach Wilseck and Paul Park and Honglak Lee and Vikas Gulani and Aditya Pandey and Todd Hollon},
  journal={arXiv preprint arXiv:2611.15094},
  year={2026}
}