A GPU-based method in recovering the full 3D deformation field using multiple 2D fluoroscopic views in lung navigation
Yixun Liu, Erkang Cheng, Henky Wibowo, Lav Rai · 2016
In this work, we present a GPU implementation for 2D-3D deformable registration to fully recover the 3D lung-deformation-field using multiple 2D fluoroscopic views. Intensive computational requirement for 2D-3D deformable registration prevents its clinical usage. To make our method clinically applicable, we developed a GPU kernel system to reduce the computation time from hours to seconds. The proposed kernel system employed a total of six GPU kernels. These six kernels serve the basis of two kernel pipelines: cost calculation and gradient calculation, the outputs of which are fed into a CPU-based optimizer to find the optimal deformation field. The evaluation was performed on both synthetic and real cases. The accuracy, evaluated using feature points, was 2.03mm on an average. A speedup factor of 690 could be reached over pure CPU-based implementation.