Fast GPU 3D diffeomorphic image registration

• Published in: Journal of parallel and distributed computing Vol. 149; no. C; pp. 149 - 162
• Main Authors: Brunn, Malte, Himthani, Naveen, Biros, George, Mehl, Miriam, Mang, Andreas
• Format: Journal Article
• Language: English
• Published: United States Elsevier Inc 01.03.2021; Elsevier
• Subjects: Computer Science; Diffeomorphic image registration; Gauss–Newton–Krylov method; GPU computing; MATHEMATICS AND COMPUTING; Mixed-precision solver; Parallel optimization; Mixed-precision solver; Diffeomorphic image registration; Parallel optimization; GPU computing; Gauss–Newton–Krylov method; Parallel Optimization; Gauss–Newton–Krylov Method; Diffeomorphic Image Registration; Mixed-Precision Solver
• ISSN: 0743-7315; 1096-0848
• DOI: 10.1016/j.jpdc.2020.11.006

3D image registration is one of the most fundamental and computationally expensive operations in medical image analysis. Here, we present a mixed-precision, Gauss–Newton–Krylov solver for diffeomorphic registration of two images. Our work extends the publicly available CLAIRE library to GPU architectures. Despite the importance of image registration, only a few implementations of large deformation diffeomorphic registration packages support GPUs. Our contributions are new algorithms to significantly reduce the run time of the two main computational kernels in CLAIRE: calculation of derivatives and scattered-data interpolation. We deploy (i) highly-optimized, mixed-precision GPU-kernels for the evaluation of scattered-data interpolation, (ii) replace Fast-Fourier-Transform (FFT)-based first-order derivatives with optimized 8th-order finite differences, and (iii) compare with state-of-the-art CPU and GPU implementations. As a highlight, we demonstrate that we can register 2563 clinical images in less than 6 s on a single NVIDIA Tesla V100. This amounts to over 20× speed-up over the current version of CLAIRE and over 30× speed-up over existing GPU implementations. •The LDDMM software CLAIRE is ported to GPU.•Compute intensive kernels are optimized.•A mixed-precision approach with Fast-Fourier-Transforms and finite differences is used.•Hardware acceleration is used for linear and cubic interpolations.•Clinical images can be registered in less than 6 seconds.