BrainIAC — Brain Imaging Adaptive Core

A generalizable foundation model for analysis of human brain MRI

BrainIAC is a Vision Transformer (ViT-B/16) pretrained with SimCLR on structural brain MRI scans. Published in Nature Neuroscience (2026).

Model Details

Property	Value
Architecture	MONAI ViT-B/16³ (3D)
Parameters	88.4M
Input	96×96×96 single-channel brain MRI
Patches	216 (6×6×6 grid, 16³ voxel patches)
Hidden dim	768
Layers	12 transformer blocks
Heads	12 attention heads
MLP dim	3072
Pretraining	SimCLR contrastive learning
Output	768-dim feature vector (first patch token)

Files

• backbone.safetensors — Pretrained ViT backbone weights
• config.json — Model configuration
• LICENSE — Non-commercial academic research license

Downstream Tasks

The backbone can be fine-tuned for:

• Brain age prediction (regression)
• IDH mutation classification (binary, dual-scan FLAIR+T1CE)
• MCI classification (binary)
• Glioma overall survival (binary, quad-scan T1+T1CE+T2+FLAIR)
• MR sequence classification (4-class: T1/T2/FLAIR/T1CE)
• Time-to-stroke prediction (regression)
• Tumor segmentation (UNETR decoder)

Usage with brainiac (Rust)

cargo run --release --bin infer -- \
    --weights backbone.safetensors \
    --input brain_t1.nii.gz

use brainiac::{BrainiacEncoder, TaskType};

let (encoder, _) = BrainiacEncoder::<B>::load(
    "backbone.safetensors", None,
    TaskType::FeatureExtraction, 1, device,
)?;
let features = encoder.encode_nifti(Path::new("brain.nii.gz"))?;
// features: Vec<f32> with 768 dimensions

Usage with Python

import torch
from monai.networks.nets import ViT
from safetensors.torch import load_file

model = ViT(in_channels=1, img_size=(96,96,96), patch_size=(16,16,16),
            hidden_size=768, mlp_dim=3072, num_layers=12, num_heads=12)

weights = load_file("backbone.safetensors")
model.load_state_dict(weights, strict=False)
model.eval()

# features[0][:, 0] gives the 768-dim feature vector
features = model(preprocessed_mri)

Preprocessing

Input MRI volumes must be:

1. Skull-stripped (HD-BET recommended)
2. Registered to standard space (MNI152)
3. Bias field corrected (N4)
4. Resized to 96×96×96 voxels (trilinear)
5. Z-score normalized (nonzero voxels only)

Citation

@article{tak2026generalizable,
  title={A generalizable foundation model for analysis of human brain MRI},
  author={Tak, Divyanshu and Gormosa, B.A. and Zapaishchykova, A. and others},
  journal={Nature Neuroscience},
  year={2026},
  publisher={Springer Nature},
  doi={10.1038/s41593-026-02202-6}
}

License

This model is licensed for non-commercial academic research use only. Commercial use requires a separate license from Mass General Brigham. See LICENSE for details.