Loading…
Brain MRI segmentation network with progressive levels of detail, trained across ~160 acquisition sites so one checkpoint handles unseen scanners.
No providers recorded yet. Browse all providers
Scan the same head on a Siemens Prisma in Glasgow and on a GE Discovery in Baltimore and the two volumes differ in intensity histogram, noise and artefacts — and a segmentation network trained at one site will quietly lose accuracy on the other. This scanner effect is why atlas-based tools such as FreeSurfer stay entrenched despite taking hours per scan: registration to a reference volume travels across sites; a voxel classifier fit to one centre's intensity statistics does not.
LOD-Brain, from Michele Svanera and Lars Muckli's group at the University of Glasgow with collaborators at the University of Brescia, attacks this by splitting the network across scales. It is a progressive level-of-detail 3D convolutional network in which each level is a U-Net seeing the same T1-weighted volume down-sampled by a further factor of four. The coarsest level sees a blurred brain where site-specific texture has been smoothed away but gross anatomy survives, so what it can learn is a scanner-independent anatomical prior: where the ventricles, cerebellum and basal ganglia sit relative to one another. Training runs bottom-up — the coarse level converges, is frozen, and its spatial context propagates upward to a finer level that refines boundaries at full resolution and takes on site-specific intensities and artefacts. The decomposition mirrors the classical atlas-then-voxel pipeline, learned end-to-end and executed in seconds.
The payoff is measured rather than asserted. Single-site models in the paper's own control experiment show a statistically significant Dice drop between held-in and held-out sites; the multi-site model shows none (p = 0.11), with no retraining on the new site. Published in Medical Image Analysis in 2024 after a 2022 preprint, it succeeds the group's CEREBRUM and CEREBRUM-7T segmenters, which were trained on single-site data and needed refitting per site.
The released configuration uses two levels with a down-sampling factor of four, taking 256³ volumes and emitting the seven classes of the MRBrainS protocol: grey matter, white matter, cerebrospinal fluid, ventricles, cerebellum, brainstem and basal ganglia. Training minimises a mixed per-channel Dice and cross-entropy loss with Adam at 5e-4 over three days on NVIDIA Quadro RTX 8000 GPUs. The corpus aggregates almost 27,000 T1w volumes from 79 datasets covering roughly 160 sites at 1.5–3T, ages 8 to 90, of which 26,169 pass MRIQC quality control; 70 datasets are internal and seven held out entirely. The model itself trains on a balanced 1,049-volume subset, 15 per dataset, because accuracy plateaus past ten to fifteen volumes per site while site count keeps mattering. Labels are FreeSurfer silver-standard masks, manually corrected only for Mindboggle101. On a balanced test set of 5,956 volumes across 77 datasets, mean Dice is 0.928 (std 0.017). Two stress tests probe generalisation: SIMON, one individual scanned in 94 sessions at 33 locations on 15 scanner models from three vendors; and five people who had undergone surgical removal of a cerebral hemisphere, where atlas-based methods hallucinate structures that are not there.
The practical target is any study pooling structural MRI across scanners or institutions — multi-centre cohorts, longitudinal work where the scanner is replaced mid-study, retrospective analyses of public repositories — where per-site harmonisation or retraining is the usual overhead. Because inference takes seconds and the model ships as Docker and Singularity images, it also fits clinical volumetry pipelines that cannot budget hours of FreeSurfer time per scan. The architecture can be retrained on a different label set.
LOD-Brain is a concrete demonstration that scanner-effect robustness can come from architecture and corpus breadth rather than from harmonisation, domain adaptation or domain randomisation, and at a hundredth the parameter count of the tools it is measured against. The limitations are worth stating plainly: it is validated for T1-weighted volumes at 1.5–3T in an 8-to-90 population and is not a 7T tool; its training targets are FreeSurfer outputs, so FreeSurfer's own errors bound what the network could learn; and seven tissue classes are coarser than a full anatomical parcellation. Code, trained weights and container images are released under an MIT licence.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.