BrainNorm

Two people can carry the same brain-age gap — five years older than their birthday, by the model's reckoning — while one is losing hippocampal volume to Alzheimer's disease and the other is losing frontal and temporal cortex to frontotemporal dementia. Collapsing a brain into a single deviation scalar cannot separate them. The usual alternative, training a generative model on healthy anatomy and reading the voxel-wise reconstruction residual, keeps the spatial detail but entangles genuine atrophy with scanner, protocol and registration noise.

BrainNorm moves the comparison out of voxel space entirely. Released as a preprint in August 2026 by researchers at IIT Hyderabad, it is a normative foundation model for 3D T1-weighted structural MRI that encodes a scan as 218 atlas-parcel embeddings in a shared Semantic Atlas Latent (SAL) space rather than as one global vector. Beside those embeddings sits an age-indexed template bank: for every parcel and every age from 10 to 100 years at 0.1-year resolution, a vector describing how that region is expected to look in a healthy brain. Asking where a subject deviates reduces to a cosine comparison, parcel by parcel, against the templates for their chronological age.

Pretraining uses healthy scans only and never sees a diagnosis. Where brain-MRI backbones such as BrainIAC, AnatCL, NeuroVFM and RadFM learn a general-purpose encoder to be fine-tuned per task, BrainNorm learns the geometry of normal aging and treats disease as distance from it.

#Key Features

• Parcel-level latent space: A shared transformer Parcel Encoder processes the token set belonging to each anatomical region independently, so anomaly scores are localized to named structures rather than to a global embedding or a raw voxel map.
• Text-seeded template bank: Templates are initialized from ~200K structured strings of the form Parcel=<name>; Tissue_class=<tissue>; Age=<a>y, embedded offline with a frozen Qwen3-8B text encoder, then reshaped by a learnable three-layer Template Mapper so they match imaging-derived aging trajectories.
• Age-aware SigLIP objective: Cross-parcel pairs act as negatives to enforce anatomical identity, while within-parcel neighbors form soft positives whose weight decays with age difference, producing smooth age-ordered trajectories instead of isolated clusters.
• Zero-shot deviation scoring: Top-K parcel deviations give a global anomaly score, and restricting the aggregation to regions implicated in a given disorder yields disease-specific scores that support label-free binary and multi-disease classification.
• Retrieval-based age estimation: Age is read off the template bank by similarity-weighted retrieval at 0.1-year precision, with the brain-age gap following from the same comparison.

#Technical Details

The Parcel Encoder is a ViT backbone over 8×8×8-voxel patches, eight layers with eight attention heads, embedding dimension 768 and feed-forward dimension 2048, projecting into a 1024-dimensional SAL space — about 67M trainable parameters. The 218-parcel atlas is a composite the authors assembled from AAL3v1 (gray matter), JHU (white-matter tracts) and Harvard–Oxford (ventricles and CSF), with all scans registered to MNI152 1mm space. Pretraining ran on roughly 38,000 healthy UK Biobank scans for 40k steps on four RTX A6000 GPUs over 100 hours, aligning ~8M parcel-template pairs. Evaluation spans 66,032 scans from 49,579 subjects across held-out UK Biobank, MCSA, ADNI, AIBL, MIRIAD, NIFD and PPMI. Direct transfer of the single UK Biobank checkpoint reaches 3.94-year age MAE on MCSA against 4.70 for the best fine-tuned baseline, and linear probes on frozen parcel-deviation features reach 94.3% AUC for AD/CN on ADNI and 80.8% for stable-versus-progressive MCI, above nine supervised baselines.

#Applications

The immediate use is individual-level screening in aging and neurodegeneration cohorts: given one T1w scan, the model returns an age estimate, a brain-age gap, and a ranked map of which parcels look atypical for that age — hippocampus and parahippocampal gyrus in Alzheimer's disease and MCI, frontal and temporal cortex in FTD, motor and subcortical structures in Parkinson's disease, matching established pathology. Because the representation is frozen and label-free, small clinical studies can attach a linear probe with as few as one to sixteen labeled examples per class rather than fine-tune a 3D network end to end, and sites can calibrate to their own scanners using healthy controls alone.

#Impact

BrainNorm reframes normative neuroimaging as a representation-learning problem, and its strongest result is that a frozen normative encoder beats end-to-end supervised training on the same preprocessed scans. The caveats are several. Under pure direct transfer, subject-level zero-shot AUC trails a classical hierarchical Bayesian regression baseline on several tasks; the authors state that they do not position BrainNorm as superior in zero-shot AUC, and its gains appear after healthy-control-only calibration on the target cohort. The pipeline depends on MNI152 registration and an adult atlas, restricting it to adult cohorts and T1w structural MRI. The work is a preprint awaiting peer review, no code or weights have been released, and pretraining on UK Biobank under Project ID 335526 would constrain any future checkpoint release.