GPU-accelerated Diffusion MRI Tractography in DIPY

Motivation: We make large-scale neuroimaging studies more practical by facilitating the generation of massive tractographies. We enable faster iteration over tractography parameters and models in smaller datasets. Goal(s): We implemented a GPU-accelerated tractography software with a myriad of methods from the Diffusion Imaging in Python (DIPY) software library, including probabilistic tractography and both single-shell/single-tissue and multi-shell/multi-tissue constrained spherical deconvolution (CSD). Approach: It is written in CUDA C, but with a python interface. It is installable using the pip package manager and mimics the DIPY API. Results: Using only a single GPU, it is up to 50x faster than equivalent CPU implementations. Impact: We introduce GPU-accelerated tractography that achieves up to 50x faster performance per GPU. Through integration with the Diffusion Imaging in Python (DIPY) library, it enables large-scale tractography studies that would not be possible otherwise.