OpenCL 기반의 고속 CT 영상 정합

초록

This paper proposes a GPU-based acceleration method to speed up 3D medical image registration using Open Computing Language (OpenCL). The initial parameters are estimated from the principal axes of binary structure of input volumes before carrying out the intensity-based optimization procedure where normalized cross correlation (NCC) is used as the objective function. Scaling factor is also considered according to the voxel size of the input volumes. The performance of the algorithm is examined on four CT patient data sets of knee image. The experimental results show that the initialization step improves the registration time algorithmically, and provides the same accuracy as that of the intensity-based approach. Moreover, the OpenCL based solution using GPUs provides a 7 to 11 speed-up factor over the CPU based implementation.