Determine whether learned priors resolve compartment degeneracy under realistic noise

Determine whether the learned priors embedded in the synthetic training data resolve the non-identifiability of multi-compartment diffusion tensors from multi-shell acquisitions with linear b-tensors under realistic in vivo noise conditions.

Background

Multi-compartment diffusion tensors are not fully identifiable from multi-shell acquisitions using linear b-tensors: multiple compartment configurations can produce identical or indistinguishable signals. The proposed network addresses this ambiguity implicitly through priors learned from synthetic data, but the paper does not establish whether those priors are sufficient under realistic in vivo noise. Stronger or more targeted priors and spatial regularization are suggested as possible ways to reduce the degeneracy.