Deformable Spherical Transformer for Cerebellar Surface Parcellation.

In neuroimaging analysis, accurate parcellation of the extremely folded cerebellar cortex is of immense importance for both structural and functional studies. To this end, we aim to develop a novel end-to-end deep learning-based method for automatic parcellation of the cerebellar cortical surface, which has an intrinsic spherical topology. Motivated by the success of Transformer, we employ Spherical Transformer to leverage its ability to model long-range dependency. To address the nonuniform, moderate distortions during the spherical mapping of the folded cerebellar surface, we propose a Deformable Spherical Transformer, which combines the Spherical Transformer architecture with the deformable attention mechanism to adaptively concentrate on the critical and challenging regions on the spherical cerebellar surface. By comparing with other state-of-the-art algorithms, we validated the superior performance of our proposed methods.