DEEP-LEARNING CORTICAL REGISTRATION GUIDED BY STRUCTURAL AND DIFFUSION MRI AND CONNECTIVITY

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Accurate cortical surface registration is vital for group neuroimaging studies.
• Existing geometry-based methods struggle with inter-individual variability, leading to suboptimal functional alignment.
• Bridging the gap between structural and functional brain organization remains a challenge.

Purpose of the Study:

• To introduce a novel deep-learning approach for improved cortical surface registration.
• To integrate white matter structural connectivity into the Joint Surface-based Registration and Atlas Construction (JOSA) framework.
• To enhance functional alignment in neuroimaging analyses by leveraging multimodal data.

Main Methods:

• Developed a deep-learning method (JOSAConn) incorporating dMRI tractography-derived white matter connectivity.
• Generated vertex-wise connectivity maps using streamline-surface intersections and heat diffusion smoothing.
• Combined connectivity features with diffusion metrics (FA, ADC) and structural data as input for JOSA.

Main Results:

• JOSAConn significantly outperformed FreeSurfer in functional alignment across 15 task contrasts (p < 0.001 for 12 contrasts) in HCP-YA subjects.
• Demonstrated superior functional alignment by integrating structural connectivity with geometric information.
• Validated the method's effectiveness on a large dataset with diverse task contrasts.

Conclusions:

• Structural connectivity effectively bridges the gap between cortical geometry and functional organization.
• The multimodal deep-learning approach enhances neuroimaging registration accuracy and functional alignment.
• This method maintains clinical applicability for advanced neuroimaging analyses.