G-Rmos: Gpu-Accelerated Riemannian Metric Optimization On Surfaces

In medical image, surface mapping plays a key role in a variety of brain imaging studies such as mapping atrophy patterns of the gray matter in Alzheimer’s disease. RMOS is a state-of-the-art surface mapping algorithm that establishes one-to-one correspondences between surfaces in the Laplace–Beltrami embedding space by optimizing the Riemannian metrics. However, RMOS takes a long time to registration due to the complex calculation with accurate establishes one-to-one correspondences. In this work, we propose GPU-accelerated RMOS registration pipeline which called G-RMOS using three strategies to CUDA kernel design: 1. utilizing computing capability of GPU with batch scheme, 2. utilizing memory hierarchy structure such that the register and shared memory, 3. hiding memory latency with ILP (Instruction Level Parallelism). Our experimental results show that G-RMOS improves the speed up to 22 times.