Spherical Transformer on Cortical Surfaces.

Area of Science:

• Neuroimaging
• Computer Vision
• Machine Learning

Background:

• Transformer architectures have shown great success in computer vision.
• Extending these models to non-Euclidean spaces like brain cortical surfaces is highly desired.
• Cortical surface data is often represented by triangular meshes with spherical topology.

Purpose of the Study:

• To propose a novel Spherical Transformer architecture for analyzing brain cortical surface data.
• To leverage the self-attention mechanism for encoding long-range dependencies and spatiotemporal patterns.
• To evaluate the model's performance on neuroimaging prediction tasks.

Main Methods:

• Mapping cortical surfaces onto a sphere and dividing them into overlapping spherical patches.
• Applying self-attention within local patches for efficient computation and contextual information preservation.
• Introducing spatiotemporal self-attention for processing longitudinal cortical surfaces and capturing dynamic patterns.

Main Results:

• The Spherical Transformer demonstrated competitive performance on surface-level (cognition prediction) and vertex-level (cortical thickness map development) prediction tasks.
• The model efficiently encodes long-range dependencies and preserves detailed contextual information.
• Spatiotemporal self-attention effectively extracts spatial context and dynamic developmental patterns.

Conclusions:

• The proposed Spherical Transformer is an effective general-purpose backbone for neuroimaging analysis.
• It offers significant improvements in processing cortical surface data, especially longitudinal data.
• The model shows promise for applications in early brain development mapping.