NL-TINR: implicit neural representation based nonlocal functional tensor decomposition for zero-shot 3D...

3D quantitative magnetic resonance imaging (qMRI) enables noninvasive tissue characterization but often requires prolonged acquisition times for multi-contrast imaging, limiting its broader clinical adoption. Although deep learning has shown promise for accelerated MRI reconstruction, supervised methods rely on large-scale fully sampled 3D datasets that are difficult to obtain. Existing implicit neural representation (INR)-based self-supervised methods avoid this requirement but can exhibit spectral bias, potentially leading to over-smoothed image details.

We propose NL-TINR, a zero-shot self-supervised reconstruction framework that integrates patch-based nonlocal self-similarity (NSS) with INR-based functional tensor decomposition. The NSS prior groups structurally similar image patches into high-dimensional tensors to explicitly exploit nonlocal redundancy and suppress aliasing artifacts while preserving fine anatomical textures. Meanwhile, INRs parameterize the tensor factor functions, providing a compact and flexible representation that captures correlations across spatial and contrast dimensions while reducing the burden of directly modeling the full 3D multi-contrast volume.

Extensive experiments on 3D multi-echo MRI datasets demonstrate that NL-TINR consistently outperforms the compared reconstruction methods, providing improved artifact suppression and anatomical detail preservation, particularly at high acceleration factors.

NL-TINR eliminates the dependence on fully sampled training data while effectively exploiting nonlocal structural redundancy and high-dimensional spatial-contrast correlations, providing a practical zero-shot framework for high-fidelity 3D multi-contrast qMRI reconstruction.