E.g., how can one sample this vast acquisition space efficiently?
Let me focus my answer on the optimization of the acquisition sampling scheme, which remains an open question (for sequence optimization, please see Szczepankiewicz et al. 2020).
To give some idea of the complexity of this issue, each analysis method will react differently to various optimization strategy. For instance, q-space trajectory imaging (QTI) is a two-term cumulant expansion and diffusion tensor distribution imaging (DTD) does not rely on such cumulant-based assumptions. As a result, QTI shows quite a low sensitivity to the details of the acquisition sampling scheme and DTD seems more sensitive to it (see Reymbaut 2021). This sensitivity should not be seen as a drawback but more as a potential for proper sampling-scheme driven performance optimization.
Within the past decade, various optimization strategies have been developed. For instance, Jespersen et al. 2013 was based on theoretical considerations, Coelho et al. 2019 optimizes precision on parameter estimation via model-specific Cramer-Rao bounds, Reymbaut 2021 optimizes the diversity in probed diffusion patterns and Tax et al. 2021 employs autoencoders in the context of machine learning.
While acquiring three b-shapes (linear, planar, spherical) can provide additional specificity, two b-shapes are usually sufficient to achieve enhanced specificity within clinically feasible times (see Nilsson et al. 2020 and Daimiel Naranjo et al. 2021). From experience, QTI and DTD provides good results for an 80-point acquisition scheme with two b-shapes (linear/spherical or linear/planar) and four b-shells: b = 100, 700, 1400, 2000 s/mm2 (~5 minutes of acquisition time). Keeping the number of points constant and adding a b-shell at b = 300 s/mm2 seems to improve the performance of DTD, as it enables better capture of the low-b signal decay.