Python GPU Accelerated Geometric Gamma Index Calculation

The gamma index is a widely adopted tool for dose distribution comparison and quality assurance in radiotherapy. The algorithm for calculating this index searches for the closest Euclidean distance, producing an easy-to-use metric for comparing dose distributions. However, this becomes computationally expensive for high-resolution images. Implementing gamma index algorithms on graphical processing units (GPUs) using C++ and MATLAB can speed up this calculation, but these existing methods are less user-friendly and modifiable, and in the case of MATLAB, expensive. This study designed a novel and efficient GPU-accelerated gamma index framework that is built on Python CUDA and Numba instead of C++. An interpolation-free geometric calculation was implemented on an NVIDIA GTX 1080 GPU with 2560 CUDA cores. The technique identifies the closest geometric distance between two distributions that can be calculated via matrix operations. The technique was modified to be GPU friendly and 3 methods were compared: non-accelerated, CUDA JIT-complied, and custom CUDA Kernel. Low-, medium-, and high-resolution image datasets were generated for benchmarking. The CUDA kernel achieved calculation speeds under 2 seconds for high-resolution 2D distributions, presenting an easy-to-modify alternative to existing GPU acceleration methods.