NeuroMorph

A 63-year-old man presents with parkinsonian signs. Structural MRI can say whether his ventricles and cerebellum sit outside the range expected for his age — but the normative charts answering that question are built one disorder at a time, each against its own matched control group. If the answer is "not typical of Parkinson's disease", nothing says what the anatomy does resemble, because progressive supranuclear palsy and corticobasal syndrome were never placed in the same coordinate system.

NeuroMorph answers that by measuring every brain the same way and placing them all in one space. It compresses a single native-space T1-weighted scan into 13 morphological descriptors, then converts that fingerprint into a deviation profile — a z-score per descriptor against age- and sex-appropriate healthy expectation. Because the normative model is fitted once over a large healthy reference cohort rather than per study, profiles from different diagnoses, cohorts and scanners are directly comparable, and an individual can be interpreted without assembling a fresh control group. It was developed at the University of Glasgow with the University of Brescia and posted as a preprint in August 2026.

Two trained networks do the measuring. LOD-Brain⁺ retrains LOD-Brain, the same group's multi-site segmentation network, to also label cerebrospinal fluid; DeepThickness reconstructs the white matter and pial surfaces and estimates cortical thickness. Both are applied as fixed checkpoints across 25 independent cohorts — the design philosophy behind the group's disease-naïve NeuroFM.

#Key Features

• Thirteen-descriptor fingerprint: Eight volumes — grey matter, white matter, intracranial, ventricular, cerebellar, brainstem, basal ganglia and CSF — plus cortical thickness, pial and white matter curvature, and pial and white matter surface area.
• No preprocessing pipeline: Scans are conformed to a 256³ 1 mm grid and intensity z-scored; there is no skull stripping, bias-field correction or MNI registration, so subject-specific geometry survives intact.
• Normative model fitted once: Bayesian linear regression over 79,674 healthy volumes, then applied unchanged to all 108,470 volumes regardless of diagnosis.
• Minutes rather than hours: Under matched conditions mean processing fell from 3.5 hours per volume with FreeSurfer's recon-all to 1.1 minutes, and to 13 seconds with a GPU.
• Hierarchical positioning: One deviation profile is read at four levels — healthy versus affected, disease family, diagnosis, and subtype.

#Technical Details

LOD-Brain⁺ is a progressive level-of-detail 3D U-Net retrained on 1,049 volumes from 11 datasets, supervised by FreeSurfer labels for seven classes and SynthSeg for the CSF mask it adds. DeepThickness is a 3D U-Net with residual bottleneck blocks, 3×3×3 kernels, group normalization and strided resampling; its decoder shares a pathway before splitting into branches that predict a level-set surface and a cortical thickness distance field, with separate checkpoints for the pial and white matter surfaces. It trained on 1,049 volumes with Adam for at most 50 epochs on an A100, and was monitored on 130 held-out cases that were 47% external. Meshes are recovered with marching cubes and decimated to 7,500 vertices. The three shipped checkpoints together hold about 1.5 million parameters.

Against FreeSurfer on 130 volumes from 10 independent datasets, mean cortical thickness differed by 0.07 mm — well under the 1 mm voxel — with NeuroMorph reporting higher grey matter and lower white matter volume, and blinded specialists preferring its pial surface placement. The reference space spans 110,591 scans from 78,794 participants aged 3–100 across 25 datasets and 19 diagnoses. Random forests on the 13-feature profiles separated healthy from affected at AUC 0.79, reached AUC 0.95 for the dementias family, and identified Parkinson's disease with 87% specificity and 58% sensitivity.

#Applications

The reference space supports three individual-level readings from one profile: ranking plausible diagnoses when presentations overlap, as with Parkinson's disease against progressive supranuclear palsy; representing comorbidity, where co-occurring autism and ADHD place a subject in the overlap region rather than at one label; and tracking drift across repeat visits, which the authors pair with Cox proportional hazards models of conversion risk. Its speed and tolerance for routine clinical acquisitions make it practical for cohort-scale morphometry and trial stratification.

#Impact

NeuroMorph reframes normative modeling from a per-disorder exercise into a shared coordinate system, and shows that a compact hand-chosen descriptor set — rather than an atlas-based parcellation of hundreds of regions — carries enough signal to organize disorders hierarchically. The authors are candid about the limits: the individual-level results are illustrative rather than confirmatory, the comorbidity finding rests on a single disease family, discrimination consistently favors exclusion over confirmation, and the diagnostic labels are clinical rather than pathologically confirmed. The 1,049 segmentation training volumes were drawn from the same aggregated cohort used in the analysis, and the work has not yet been peer reviewed. A public inference container carrying both networks' weights is available; the source and weight repositories named in the paper are not yet public.