STF: A spherical transformer for versatile cortical surfaces applications.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Imaging Analysis

Background:

• Transformer architectures excel in Euclidean domains but struggle with non-Euclidean medical imaging data.
• Brain cortical surfaces, with their spherical topology, require specialized analysis methods.
• Existing methods may not fully capture both global context and fine-grained details in surface data.

Purpose of the Study:

• To adapt the Transformer architecture for analyzing non-Euclidean medical imaging data, specifically brain cortical surfaces.
• To develop a versatile backbone model, the Spherical Transformer (STF), for enhanced cortical surface analysis.
• To introduce a spatiotemporal self-attention mechanism for modeling longitudinal surface data.

Main Methods:

• Mapping cortical surfaces onto a sphere and dividing them into overlapping patches.
• Tokenizing patches and vertices for self-attention at both levels.
• Implementing a spatiotemporal self-attention mechanism for longitudinal data analysis.

Main Results:

• The Spherical Transformer (STF) effectively models global dependencies and preserves fine-grained spatial information.
• STF consistently outperformed state-of-the-art methods on cognition prediction, cortical surface parcellation, and property map prediction.
• The overlapping patch design facilitates efficient cross-patch information sharing.

Conclusions:

• The Spherical Transformer (STF) is a powerful and versatile framework for analyzing cortical surface data.
• The model's ability to handle both spatial and temporal dynamics makes it suitable for longitudinal studies.
• STF demonstrates significant potential for advancing medical imaging analysis and understanding brain structure and function.