GPU-accelerated spherical-deconvolution informed filtering of tractograms for quantitative diffusion MRI connectomics

Streamline tractography from diffusion magnetic resonance imaging is the standard technique for mapping the structural connections of the human brain, but raw tractograms contain systematic streamline-density biases that undermine quantitative connectivity analysis. Spherical-deconvolution informed filtering of tractograms (SIFT) and its successor SIFT2 correct these biases by matching streamline densities to the underlying fibre orientation distributions; however, their computational cost-from tens of minutes to several hours for a 10-million-streamline tractogram-restricts their use in large cohort studies. We describe the design, implementation and validation of graphics processing unit (GPU) implementations of both algorithms using the CUDA programming model. A two-stage architecture retains the reference CPU implementation for streamline-to-fixel mapping and offloads only the optimisation phase to the GPU, so that the accelerated code can be validated directly against the reference software on identical inputs. On tractograms of 10 million streamlines from four subjects drawn from the UK Biobank and the Human Connectome Project, the GPU SIFT implementation ran 25 to 56 times faster than the multi-threaded CPU reference while selecting an essentially identical streamline subset (at most 30 of 10 6 retained streamlines differed per subject; overlap at least 99.997%). The GPU SIFT2 implementation ran 6.8 to 9.3 times faster with a coefficient of determination of at least 0.999998 between GPU and CPU streamline weights. Both implementations accept unmodified MRtrix3 inputs and produce outputs in the same formats as the reference commands, allowing direct substitution into existing tractography pipelines.