Single- and Multiple-Shell Uniform Sampling Schemes for Diffusion MRI Using Spherical Codes

In diffusion MRI (dMRI), a good sampling scheme is important for efficient acquisition and robust reconstruction. Diffusion weighted signal is normally acquired on single or multiple shells in q-space. Signal samples are typically distributed uniformly on different shells to make them invariant to the orientation of structures within tissue, or the laboratory coordinate frame. The Electrostatic Energy Minimization (EEM) method, originally proposed for single shell sampling scheme in dMRI, was recently generalized to multi-shell schemes, called Generalized EEM (GEEM). GEEM has been successfully used in the Human Connectome Project (HCP). However, EEM does not directly address the goal of optimal sampling, i.e., achieving large angular separation between sampling points. In this paper, we propose a more natural formulation, called Spherical Code (SC), to directly maximize the minimal angle between different samples in single or multiple shells. We consider not only continuous problems to design single or multiple shell sampling schemes, but also discrete problems to uniformly extract sub-sampled schemes from an existing single or multiple shell scheme, and to order samples in an existing scheme. We propose five algorithms to solve the above problems, including an incremental SC (ISC), a sophisticated greedy algorithm called Iterative Maximum Overlap Construction (IMOC), an 1-Opt greedy method, a Mixed Integer Linear Programming (MILP) method, and a Constrained Non-Linear Optimization (CNLO) method. To our knowledge, this is the first work to use the SC formulation for single or multiple shell sampling schemes in dMRI. Experimental results indicate that SC methods obtain larger angular separation and better rotational invariance than the state-of-the-art EEM and GEEM. The related codes and a tutorial have been released in DMRITool.

💡 Research Summary

This paper addresses the fundamental problem of designing efficient q‑space sampling schemes for diffusion MRI (dMRI). While the widely adopted Electrostatic Energy Minimization (EEM) and its multi‑shell extension Generalized EEM (GEEM) generate uniformly distributed gradient directions, they optimize a surrogate objective—the total electrostatic energy of all point pairs—rather than directly maximizing the angular separation that is most relevant for dMRI reconstruction. The authors therefore propose a more natural formulation based on Spherical Codes (SC), which seeks to maximize the covering radius, i.e., the smallest angular distance between any two sampling points on the unit sphere.

The SC framework is first defined for a single shell: given K points {u_i} on S², the covering radius d = min_{i≠j} arccos|u_iᵀu_j| is maximized. This leads to two problem classes: continuous SC (CSC), where the points can be placed anywhere on the sphere, and discrete SC (DSC), where a finite candidate set of N points is given and K must be selected. The authors extend the formulation to multiple shells by maximizing a weighted mean of the covering radii of each individual shell and of the combined set of all shells; the weight w is typically set to zero so that inter‑shell staggering is encouraged.

Five algorithms are introduced to solve the various SC‑based design tasks:

1. Incremental SC (ISC) – adds points sequentially while preserving the current covering radius; useful for the acquisition ordering problem (P‑O) because it requires no initialization.
2. Iterative Maximum Overlap Construction (IMOC) – a greedy procedure that, at each iteration, selects the candidate that yields the largest increase in the covering radius. It works without any prior solution and provides a strong baseline for both continuous and discrete problems.
3. 1‑Opt – a local refinement that swaps a single point in an existing configuration to improve the covering radius; typically applied after IMOC.
4. Mixed‑Integer Linear Programming (MILP) – formulates the discrete selection problem as an integer linear program, allowing near‑global optimality for moderate‑size candidate sets. It is especially effective for the “multiple subsets from a single set” (P‑D‑MS) and “multiple subsets from multiple sets” (P‑D‑MM) cases.
5. Constrained Non‑Linear Optimization (CNLO) – directly tackles the continuous SC problem with non‑linear constraints; it yields the highest-quality solutions but at a substantial computational cost.

The authors categorize the design tasks into six problem types: single‑shell continuous (P‑C‑S), multi‑shell continuous (P‑C‑M), single‑set sub‑sampling (P‑D‑SS), multi‑set sub‑sampling (P‑D‑MM), multi‑subset from a single set (P‑D‑MS), and acquisition ordering (P‑O). Table I in the paper maps each algorithm to the problems it can address, and Table II recommends specific algorithmic pipelines (e.g., IMOC + 1‑Opt + CNLO for continuous problems, MILP for discrete problems).

Experimental validation includes synthetic data with random rotations and added noise, as well as real Human Connectome Project (HCP) multi‑shell acquisitions (b‑values 1000, 2000, 3000 s/mm²). Results consistently show that SC‑based schemes achieve larger covering radii—typically 5–10° higher than EEM/GEEM—leading to better rotational invariance and lower reconstruction error across a range of diffusion models (DTI, HARDI, ODF reconstruction). The multi‑shell designs also exhibit reduced inter‑shell redundancy, which improves angular resolution without increasing scan time. Computationally, IMOC runs in O(K·N) time and scales well to hundreds of directions; MILP provides exact solutions for N < 200 but becomes prohibitive for larger N; CNLO offers the best quality but requires careful initialization (often from IMOC + 1‑Opt).

In summary, the paper introduces the spherical‑code formulation as a principled alternative to electrostatic‑energy‑based designs for diffusion MRI sampling. By directly maximizing the minimal angular separation, the proposed methods deliver superior uniformity, enhanced rotational robustness, and improved reconstruction fidelity. The work also supplies a comprehensive algorithmic toolbox—ISC, IMOC, 1‑Opt, MILP, and CNLO—covering both continuous and discrete design scenarios, and makes the code publicly available through the DMRITool package. This contribution bridges a gap between mathematical coding theory and practical dMRI acquisition, offering a new standard for future gradient‑direction scheme development.