Regularized Joint Reconstruction and Slab Combination for Accelerated Three-Dimensional Multi-Slab Diffusion-Weighted Imaging Using Multi-Scale Energy Models.

purposeTo jointly reconstruct high-resolution diffusion-weighted volumes and eliminate slab-boundary artifacts while preserving fine anatomical detail from undersampled 3D multi-slab k-space acquisitions.

methodsA bilinear forward model was formulated to describe the 3D multi-slab acquisition, treating the image volume and slab excitation profiles as joint variables. A maximum a posteriori framework optimized a cost function with a Gaussian data-fidelity term, a CNN-based deep energy prior guiding accelerated recovery, and a quadratic regularization term anchoring the slab profiles to an initial estimate. The energy-based prior models the negative log distribution of clean diffusion-weighted images, while its gradient guides the accelerated reconstruction toward the artifact-free distribution. The resulting non-convex problem was solved via alternating minimization, where the image volume was updated through a majorize-minimize scheme with conjugate gradient optimization, and slab profiles were estimated through regularized least-squares optimization.

resultsThe proposed Energy-based Profile Encoding (EPEN) framework substantially reduced slab-boundary artifacts compared with conventional slab-boundary correction approaches. Improved structural consistency and contrast preservation were achieved across multiple acceleration factors and slab configurations.

conclusionEPEN enables robust joint 3D multi-slab diffusion MRI reconstruction with slab profile correction within an alternating optimization framework using deep energy-based image priors.