8001 Single and Multiple Shell Sampling Design in dMRI Using Spherical code and Mixed Integer Linear Programming
Jian Cheng,Pew-Thian Yap,Dinggang Shen
2013-01-01
Abstract:Introduction. Sampling scheme is crucial in diffusion MRI data acquisition and reconstruction. A good sampling scheme can obtain good reconstruction results with less number of samples in diffusion q-space. It has been showed that for the diffusion data obtained from single shell (single b value), the uniform sampling scheme has good performance in a general case because it does not assume a preferred direction. In last decade, two kinds of uniform single shell sampling schemes are widely used in diffusion MRI field, i.e. the sphere tessellation and electrostatic energy minimization introduced in dMRI by Jones . However sphere tessellation cannot handle arbitrary number of samples, and electrostatic energy minimization lacks its physical meaning related to reconstructions in dMRI. It is still unclear that why electrostatic energy matters in diffusion data reconstruction? We hope the samples in sphere have large angular difference such that the “angular resolution” in reconstruction can be as large as possible, so mathematically a more natural way is to define a uniform sampling scheme {ui} such that the minimal angular difference can be as large as possible, which is essentially the Spherical Code (SC) problem. Some optimal configurations for sphere S for different number K have been collected in [3]. Although the solutions in [3] can be directly used in dMRI, they have two main limitations. 1) They are for single sphere, not for multi-shell schemes. 2) For real applications in dMRI, sometimes we need to find one or several sets of “uniform” schemes from a given set of sampling scheme, which is the “discretized” SC problem. In this paper, we propose a general Mixed Integer Linear Programming (MILP) framework to design “uniform” sampling schemes for both single and multi-shell dMRI. Although some recent works proposed to generalized electrostatic energy minimization method to multi-shell case , to our knowledge, this paper is the first work to design uniform single/multi-shell sampling scheme using SC formulation. Theory: Single shell case. Given the number K, the SC problem is to find the K samples in single shell {ui} such that the minimal distance is maximized. See Eq (1). The absolute value is used because in dMRI we want to force the antipodal symmetric constraint. Note that the original SC problem assumes the search domain D is the continuous sphere S. In dMRI we may need to search the solution from a given set of samples with size N. We formulate this problem as a mixed integer linear programming as Eq (2), where hi =1 means i-th sample is chosen, dub (2K) is the theoretical upper bound for 2K samples given in [1], dlb is an lower bound which can be chosen by some prior knowledge or any feasible solution. M is the difference between the maximal distance and the minimal distance of any two pair ui and uj , such that the inequality constraint holds if and only if both i-th and j-th samples are chosen. This problem can be solved by GUROBI or other software. Please note that the obtained solution of this discretized SC problem can be used as an initialization of the continuous SC problem. Any local search method can be used to improve the solution in the continuous domain. Theory: Multi-shell case. The above discretized SC method for single shell can be generalized to multi-shell case, by defining the cost function as Eq. (3), where there are Ns shells, and every shell has Ks different directions from the given ∑sKs directions. By considering all directions in different single shells as a whole shell, we can maximize the angular resolution between shells, thus w is the compromise weight between the single shell term and the multi-shell term. Similarly with Eq (1), Eq (3) also can be transformed to a mixed integer linear programming problem in Eq (4), where hs,i.=1 means direction ui is chosen by s-th shell. Compared to Eq (2), Eq(4) incorporates some constrains to avoid ambiguity between shells. Note that although the solution of Eq (4) is from a given discrete set of directions, it can be used as an initialization to obtain better solution in continuous multi-shells. The formulation (4) can be also used for single shell scheme to separate a given single shell sampling scheme into several parts, as the tool subsetpoints in CAMINO [5], which is very slow because the simulated-annealing is used. If w=1, these separated parts have no angular exclusion.