Accelerated Cortical Thickness Mapping Using Deep Learning

Motivation: The long processing time of current CT mapping methods hampers their use in clinical research. A faster and reliable CT mapping alternative is needed. Goal(s): To create a deep-learning model that reduces CT mapping time without compromising accuracy or the ability to classify Alzheimer's disease. Approach: We trained a 3D U-Net-based model on T1-weighted MRI datasets to produce CT maps, generating two model variants—one using skull-stripped and the other using both whole-brain and skull-stripped images. Performance was benchmarked against FreeSurfer. Results: The complete Unet-based CT mapping workflow, inclusive of preprocessing, was executed in under a minute without relying on GPU acceleration. Impact: The developed deep-learning-based method, executed within a minute, could accelerate neurological research related to CT values by providing fast and reliable procedure for CT mapping.