SphereMorph

SphereMorph performs diffeomorphic registration of cortical surfaces using a UNet-style convolutional network and spherical kernels to learn displacement fields for accurate inter-subject alignment of cortical folding and functional areas.

Key Features:

• Diffeomorphic registration framework: Implements a non-rigid diffeomorphic registration approach tailored to cortical surface geometry.
• UNet-style CNN: Uses a UNet-style deep convolutional network to learn displacement fields for surface alignment.
• Spherical kernel: Employs a spherical kernel to operate on spherical representations of cortical surfaces.
• Modified spatial transformer layer: Applies a modified spatial transformer layer to warp the spherical surface according to the learned displacement field.
• Resampling-weighted data-fitting loss: Introduces a resampling weight in the data-fitting loss to compensate for distortions from polar projection.
• Geometric registration in CNN framework: Frames cortical surface registration as a geometric registration problem solvable within a CNN architecture.
• Reported performance: Experimental results report superior registration accuracy and computational efficiency compared to traditional spherical registration methods.

Scientific Applications:

• Cortical parcellation: Applied and evaluated on cortical parcellation tasks to align anatomical boundaries across subjects.
• Group-wise functional area alignment: Tested for group-wise alignment of functional areas on spherical cortical representations.
• Non-rigid cortical registration: Targets non-rigid alignment of cortical folding patterns across individuals.

Methodology:

Trains a UNet-style CNN with spherical kernels to learn displacement fields on spherical cortical representations, uses a modified spatial transformer layer to warp the sphere, and incorporates a resampling weight in the data-fitting loss to correct polar projection distortions.