Continuous q-Space Signal Synthesis from Arbitrary Sparse Multi-Shell Measurements

Dense multi-shell diffusion MRI acquisitions provide rich q-space information but require long scan times, limiting their practicality in clinical and large-scale settings. We propose a spatially aware teacher-student framework for reconstructing continuous q-space information from sparse multi-shell observations. The method treats the observed diffusion measurements as an unordered masked set and predicts SHORE coefficients for the center voxel from a lightweight \(3\times3\times3\) neighborhood. Architecturally, the model combines shared per-measurement encoding, gated local-and-wide spatial attention, cross-gradient self-attention, and permutation-invariant pooling, yielding a structured latent representation that can be decoded analytically to synthesize signals at arbitrary q-space locations. To improve robustness under sparse acquisition, we train a teacher on full observations and distill its pooled latent representation into sparse-input student model using an MSE-based feature matching loss. We further employ shell-wise antipodal repulsive sampling and SHORE-consistent rotation augmentation to simulate realistic sparse protocols while preserving physical consistency between inputs and supervision. Experiments on held-out Human Connectome Project subjects show that the proposed method consistently outperforms analytical baselines and recent learning-based reconstruction methods across sparse single-shell and multi-shell settings, with the largest gains in the low-measurement regime. These results demonstrate that combining set-based q-space modeling, local spatial context, and teacher-guided distillation enables accurate and flexible diffusion signal reconstruction from limited observations. Our implementation is available on \href{https://github.com/xmindflow/SHOREPred}{GitHub}.