Hierarchical Mesh Transformers with Topology-Guided Pretraining for Morphometric Analysis of Brain Structures

A research team led by Yujian Xiong has developed a novel AI architecture called the Hierarchical Mesh Transformer, designed to overcome a major hurdle in computational neuroimaging. Current methods are typically limited to analyzing either 3D volumetric brain scans or 2D cortical surface meshes, but not both within a single model. This new framework operates on spatially adaptive tree partitions built from 'simplicial complexes,' a mathematical structure that allows it to handle both types of brain data representations seamlessly. Crucially, it incorporates a feature projection module that maps diverse clinical measurements—like cortical thickness, curvature, and myelin content—directly into the model's spatial hierarchy, separating geometric structure from feature dimensionality.

The model was pretrained in a self-supervised manner using masked reconstruction on large, unlabeled datasets, creating a transferable encoder backbone. This pretraining approach allows the model to learn robust representations without needing vast amounts of hard-to-obtain labeled medical data. The team validated their system on major neuroimaging benchmarks: Alzheimer's disease classification and amyloid burden prediction using volumetric meshes from the ADNI dataset, and focal cortical dysplasia detection on surface meshes from the MELD dataset. The Hierarchical Mesh Transformer achieved state-of-the-art results across all these tasks, demonstrating its versatility and superior performance compared to previous topology-specific models.