NeuroFM

In the authors' own preliminary analyses, a standard 3D residual CNN and a model built from nothing but MRIQC image-quality metrics each separated frontotemporal dementia from healthy controls at AUC above 0.85. Neither had learned anything about atrophy — they had learned which scanner produced which scan. That is the recurring failure of representation learning on labelled clinical neuroimaging: the model absorbs whichever cue predicts the label, and cohort composition, acquisition protocol and disease effect are entangled in every real dataset.

NeuroFM removes the labels entirely. Developed in the School of Psychology & Neuroscience at the University of Glasgow and posted as a medRxiv preprint in March 2026, it is pretrained exclusively on 100,000 AI-generated healthy brain volumes from the LDM100k corpus — synthetic T1-weighted MRI produced by a latent diffusion generator fit to UK Biobank anatomy — supervised only by four interpretable targets: chronological age, biological sex, total brain volume, and lateral ventricular volume. No real scan and no diagnosis enters pretraining. The resulting checkpoint is then frozen, and every downstream result comes from a linear or logistic readout on its features.

This inverts the usual arrangement. Where BrainIAC learns from unlabelled real clinical MRI through contrastive self-supervision, NeuroFM treats population-scale synthetic anatomy as its only teacher and morphometry as a proxy that forces the network to represent global and regional shape. The evaluation corpus — 136,361 real multi-cohort volumes — is disjoint from pretraining by construction.

#Key Features

• Disease-naive pretraining: The model never sees a pathological volume or a diagnostic label, so its representation cannot encode cohort- or disease-specific shortcuts. Diagnostic structure emerges from anatomy alone.
• Frozen deployment: Downstream tasks use fixed parameters with simple linear probes, cross-validated with folds grouped by individual, rather than per-cohort refitting.
• Two inference modes: Predictor mode returns the four brain-health markers end to end; encoder mode returns a 161-, 256-, or 512-dimensional embedding from the penultimate layer for classification, regression, or clustering.
• Parameter-efficient at three scales: Bottleneck compression blocks keep the S, M, and L variants to 484,829, 6,535,877, and 10,815,109 parameters; the smallest runs comfortably on CPU.
• Spatial attribution: Layer-wise relevance propagation yields anatomically coherent maps that match established FTD subtype profiles and localise gliomas in BraTS.

#Technical Details

NeuroFM is a 3D convolutional encoder, not a transformer: bottleneck residual blocks with strided-convolution downsampling feed global average pooling, fully connected neck layers, and four separate prediction heads. Inputs are skull-stripped, 1 mm isotropic volumes conformed to a 256³ grid in native space, with no spatial registration. Training ran for 60 epochs in TensorFlow 2.13 under Adam, with geometric, intensity, and artefact augmentations applied to 95% of volumes. The M variant achieved the best pretraining fit (R² = 0.86 for age, 0.97 for ventricular volume). On frozen features, NeuroFM reached AUC 0.83 for FTD versus controls in NIFD, against 0.75 for BrainIAC, 0.69 for SFCN, and 0.58 for a matched ResNet-style CNN; 0.77 for Alzheimer's disease versus controls in ADNI; and 0.72 for glioma size classification in BraTS 2023. Neurodevelopmental signals were weak but reproducible, at AUC 0.59 for both ADHD-200 and ABIDE. Linear regressions on the same features recovered MRIQC quality metrics — contrast-to-noise MAE 0.49 — in under ten seconds per volume without template registration. Because acquisition artefacts can masquerade as pathology, all downstream discriminative analyses were confound-corrected with double machine learning against quality-control variables.

#Applications

The model targets cohort-scale neuroimaging where per-study training is impractical: brain-age gap as a biomarker, differential diagnosis from structural MRI, template-free quality control on volumes that registration pipelines fail, and individualised risk profiling. In ADNI, normative trajectories from cognitively normal participants plus a Cox proportional hazards model on brain age gap flagged elevated conversion risk to Alzheimer's disease up to five years before diagnosis, with converters averaging 1.52 years higher gap over the pre-conversion window. Fine-tuning extends it beyond its pretraining contrast: 500–1,000 T2-weighted UK Biobank volumes approached the T1-weighted model's age-prediction accuracy, where training the same architecture from scratch needed at least 40,000.

#Impact

NeuroFM's contribution is evidence that a foundation model for brain health can be built without touching patient data at all — sidestepping consent and governance constraints while avoiding the age-sampling bias of real cohorts, since LDM100k is uniformly distributed across age. That the frozen representation then outperforms models pretrained on the very cohorts being evaluated, including BrainIAC on ABIDE and ADNI, argues that anatomy-grounded proxy supervision generalises better than task-specific optimisation. The scope is narrow: the work is a preprint awaiting peer review, the model handles only cross-sectional structural MRI, its neurodevelopmental performance sits close to chance, and the authors state it is for research use and is not validated for clinical decisions. Code and all three weight variants are released under CC BY-NC-SA 4.0, which permits research reuse but not commercial deployment; the Hugging Face repository gates every file download behind an automatically approved contact-information agreement.