ScaleSurfer

Quantitative brain morphometry—measuring regional volumes, cortical thickness, surface area, and curvature from structural MRI—underpins a large fraction of neuroscience and clinical neuroimaging, from tracking neurodegeneration to studying development and individual differences. The established way to obtain these measures relies on surface-reconstruction pipelines that segment tissue, build cortical surfaces, and parcellate the brain. These pipelines are accurate and widely trusted, but they are computationally heavy, typically requiring hours of processing per scan, which becomes a bottleneck for large cohorts and rules out interactive or point-of-care use.

ScaleSurfer, developed by Ryan P. Hammonds, Cindy Chen, and Bradley Voytek at the University of California, San Diego and its Halıcıoğlu Data Science Institute, reframes morphometry as a direct prediction problem. Rather than reconstructing surfaces, the model maps a structural MRI volume straight to a full set of whole-brain morphometric features using a 3D convolutional vision-transformer backbone. This collapses a multi-hour workflow into a single forward pass, estimating features in roughly 150–500 milliseconds—about five orders of magnitude faster than classical surface-based processing—while retaining the multi-scale, region-resolved output that downstream analyses depend on.

The model is trained across multiple heterogeneous neuroimaging collections so that a single pretrained backbone generalizes across scanners, acquisition protocols, and populations. That shared representation is reused beyond feature estimation: the authors build an interpretable Alzheimer's disease classifier on top of the same backbone, illustrating the foundation-model pattern of pretrain-once, adapt-to-many.

#Key Features

• Direct morphometry from MRI: Predicts volume, cortical thickness, surface area, and curvature for the whole brain without running surface reconstruction or parcellation.
• Millisecond inference: Produces a complete morphometric profile in roughly 150–500 ms per scan, versus the multi-hour runtime of conventional pipelines.
• Multi-scale anatomical output: Resolves features at multiple anatomical scales, preserving the regional detail needed for downstream statistical and clinical analysis.
• Convolutional vision-transformer backbone: Combines 3D convolutions with transformer attention to capture both local tissue structure and long-range anatomical context.
• Reusable pretrained representation: The backbone, pretrained across diverse datasets, transfers to downstream tasks such as an interpretable Alzheimer's disease classifier.

#Technical Details

ScaleSurfer is a 3D model that pairs convolutional feature extraction with transformer self-attention to regress morphometric features directly from volumetric structural MRI. It is pretrained across multiple heterogeneous neuroimaging datasets spanning different scanners and populations, which supports generalization to inputs outside any single acquisition protocol. Reported inference time is approximately 150–500 milliseconds per scan on GPU hardware, compared with the several hours required by classical surface-reconstruction morphometry, corresponding to an improvement of roughly five orders of magnitude in speed. Beyond regression of morphometric features, the same pretrained backbone is adapted into a downstream classifier for Alzheimer's disease that is designed to remain interpretable in terms of the morphometric features it draws on.

#Applications

ScaleSurfer targets neuroimaging researchers and clinical scientists who need morphometric measurements at scale or in near real time. Its speed makes it practical to process very large cohorts, to iterate on quality control interactively, and to integrate morphometry into settings where multi-hour batch processing is infeasible. Because the model outputs the same families of features—thickness, area, volume, curvature—used throughout the morphometry literature, it can slot into existing analysis workflows for studies of development, aging, and neurodegeneration, while the interpretable Alzheimer's classifier demonstrates how the shared backbone can support disease-focused applications.

#Impact

By turning a multi-hour surface-reconstruction pipeline into a sub-second inference step, ScaleSurfer addresses one of the most persistent throughput bottlenecks in structural neuroimaging and lowers the computational barrier to population-scale morphometry. Its foundation-model framing—pretraining a single backbone across heterogeneous data and reusing it for both feature estimation and downstream disease classification—aligns with the broader shift toward reusable, adaptable models in medical imaging. As a preprint awaiting peer review, its measurements should be understood as fast approximations to established pipeline outputs, and, as with any learning-based estimator, accuracy depends on inputs resembling the distribution of brain MRI it was trained on.