Evaluation of volume-based and surface-based brain image registration methods.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Image Analysis

Background:

• Brain image registration is crucial for comparative studies and group analysis.
• Existing evaluations of automated nonlinear registration methods have primarily focused on volume-based techniques.
• A gap exists in comparing advanced volume-based methods with surface-based registration strategies.

Purpose of the Study:

• To conduct the first direct comparison of leading volume-based and surface-based brain image registration methods.
• To evaluate the impact of de-skulling (brain-only extraction) versus whole-head registration.
• To assess the performance of custom-made average templates versus direct pairwise registration.

Main Methods:

• Utilized permutation tests to compare registration performance (overlap, Hausdorff distance) across over 16,000 registrations.
• Included 80 manually labeled brain images for comprehensive evaluation.
• Compared all combinations of volume-based and surface-based labels, registration algorithms, and evaluation metrics.

Main Results:

• De-skulling significantly enhances the performance of volume registration methods.
• Custom-designed optimal average templates yield superior registration results compared to direct pairwise registration.
• Resampling volume labels to surfaces or vice-versa introduces distortions, preventing fair comparison between top-tier volume and surface methods.

Conclusions:

• De-skulling is beneficial for improving volume-based brain image registration.
• Employing custom average templates constructed from representative samples improves registration accuracy.
• Current resampling techniques limit the direct comparison of advanced volume and surface registration methods, necessitating careful consideration for future studies.