Optimized Diffusion Imaging for Brain Structural Connectome Analysis.

High angular resolution diffusion imaging (HARDI), a type of diffusion magnetic resonance imaging (dMRI) that measures diffusion signals on a sphere in q-space, is widely used in data acquisition for human brain structural connectome analysis. Accurate estimation of the local diffusion, and thus the structural connectome, typically requires dense sampling in HARDI, resulting in long acquisition times and logistical challenges. We propose a method to select an optimal set of q-space directions for recovery of the local diffusion under a sparsity constraint on the sampling budget. Relevant historical dMRI data is leveraged to estimate a prior distribution of the local diffusion in a template space using reduced rank Gaussian process models. For a new subject to be scanned, the priors are mapped into the subject-specific coordinate and used to guide an optimized q-space sampling which minimizes the expected integrated squared error of a diffusion function estimator from sparse samples. The optimized sampling locations are inferred by an efficient greedy algorithm with theoretical bounds approximating the global optimum. Simulation studies and a real data application using the Human Connectome Project data demonstrate that our proposed method provides substantial advantages over its competitors.