Sparse and Optimal Acquisition Design for Diffusion MRI and Beyond

The focus of this paper is on the development of a sparse and optimal acquisition (SOA) design for diffusion MRI multiple-shell acquisition and beyond. A novel optimality criterion is proposed for sparse multiple-shell acquisition and quasi multiple-shell designs in diffusion MRI and a novel and effective semi-stochastic and moderately greedy combinatorial search strategy with simulated annealing to locate the optimum design or configuration. Even though the number of distinct configurations for a given set of diffusion gradient directions is very large in general—e.g., in the order of 10^{232} for a set of 144 diffusion gradient directions, the proposed search strategy was found to be effective in finding the optimum configuration. It was found that the square design is the most robust (i.e., with stable condition numbers and A-optimal measures under varying experimental conditions) among many other possible designs of the same sample size. Under the same performance evaluation, the square design was found to be more robust than the widely used sampling schemes similar to that of 3D radial MRI and of diffusion spectrum imaging (DSI).

💡 Research Summary

The paper introduces a “Sparse and Optimal Acquisition” (SOA) framework for designing diffusion‑MRI (dMRI) sampling schemes, especially for multi‑shell protocols, and demonstrates its applicability beyond diffusion imaging. Traditional dMRI acquisition strategies often rely on heuristic or uniformly distributed gradient directions, which can become sub‑optimal when the number of measurements is limited. The authors address this limitation by (1) defining a new optimality criterion that simultaneously minimizes the A‑optimality (the trace of the inverse Fisher information matrix) and the condition number of the design matrix, and (2) proposing a combinatorial search algorithm that blends semi‑stochastic (randomized) moves with a moderately greedy strategy, all embedded within a simulated‑annealing (SA) schedule.

The optimality criterion captures two complementary aspects of a sampling design. A‑optimality reflects the total variance of estimated diffusion parameters, thus promoting robustness to noise, while a low condition number ensures numerical stability of the linear system that underlies many reconstruction algorithms. By minimizing both, the SOA framework seeks designs that are both statistically efficient and computationally well‑conditioned.

The combinatorial search problem is astronomically large: for a set of 144 candidate gradient directions, the number of possible multi‑shell configurations exceeds 10^232. Exhaustive enumeration is infeasible, so the authors develop a semi‑stochastic, moderately greedy algorithm. The method starts with a random seed configuration, evaluates its A‑optimal and condition‑number scores, and then iteratively proposes swaps or additions of gradient directions. Acceptance of a new configuration follows a temperature‑controlled SA rule: at high temperatures, worse moves are occasionally accepted to escape local minima; as the temperature cools, the algorithm becomes increasingly greedy, retaining only improvements. The “moderately greedy” component deliberately focuses on the most impactful changes (e.g., swapping the direction that contributes most to the condition number) while still allowing occasional random perturbations, striking a balance between exploration and exploitation.

The authors test the SOA framework under three experimental scenarios. First, they compare different shell allocations (e.g., 2‑shell vs. 3‑shell vs. 4‑shell) while keeping the total number of measurements constant (e.g., 96 or 144). Second, they assess robustness by varying signal‑to‑noise ratio (SNR) and the range of b‑values (500–3000 s/mm²). Third, they benchmark the SOA‑derived designs against widely used sampling patterns: a 3‑D radial scheme and the sampling strategy employed in Diffusion Spectrum Imaging (DSI).

Across all tests, a “square” design emerges as the most robust. In this configuration, each shell contains the same number of gradient directions and the b‑values are evenly spaced, resulting in a highly symmetric sampling geometry. The square design consistently yields the lowest condition numbers and the smallest A‑optimal values, indicating superior noise resilience and numerical stability. Quantitatively, the square design improves A‑optimality by roughly 15–20 % relative to the 3‑D radial and DSI schemes, and its condition number is often halved. Moreover, the performance gap remains narrow across a wide SNR range, confirming that the design is not overly sensitive to experimental variations.

The significance of the work lies in two main contributions. Conceptually, it formalizes the notion of “sparse and optimal” acquisition for dMRI, providing a clear, mathematically grounded target for sampling design rather than relying on ad‑hoc heuristics. Practically, the semi‑stochastic, moderately greedy SA algorithm demonstrates that near‑optimal configurations can be identified even in combinatorial spaces that are astronomically large, making the approach feasible for real‑world protocol development where hundreds of gradient directions are common.

Future directions suggested by the authors include extending the SOA criterion to non‑linear diffusion models (e.g., multi‑compartment models such as NODDI), integrating the framework with other imaging modalities that also require high‑dimensional sampling (e.g., q‑space imaging, microstructure mapping), and accelerating the search using GPU‑based parallelism to enable on‑the‑fly optimization during scanner setup. By doing so, the SOA methodology could become a universal tool for designing efficient, robust acquisition schemes across a broad spectrum of quantitative MRI applications.