From Coarse to Continuous: Progressive Refinement Implicit Neural Representation for Motion-Robust Anisotropic MRI Reconstruction

Abstract

In motion-robust magnetic resonance imaging (MRI), slice-to-volume reconstruction is critical for recovering anatomically consistent 3D brain volumes from 2D slices, especially under accelerated acquisitions or patient motion. However, this task remains challenging due to hierarchical structural disruptions. It includes local detail loss from k-space undersampling, global structural aliasing caused by motion, and volumetric anisotropy. Therefore, we propose a progressive refinement implicit neural representation (PR-INR) framework. Our PR-INR unifies motion correction, structural refinement, and volumetric synthesis within a geometry-aware coordinate space. Specifically, a motion-aware diffusion module is first employed to generate coarse volumetric reconstructions that suppress motion artifacts and preserve global anatomical structures. Then, we introduce an implicit detail restoration module that performs residual refinement by aligning spatial coordinates with visual features. It corrects local structures and enhances boundary precision. Further, a voxel continuous-aware representation module represents the image as a continuous function over 3D coordinates. It enables accurate inter-slice completion and high-frequency detail recovery. We evaluate PR-INR on five public MRI datasets under various motion conditions (3% and 5% displacement), undersampling rates (4x and 8x) and slice resolutions (scale = 5). Experimental results demonstrate that PR-INR outperforms state-of-the-art methods in both quantitative reconstruction metrics and visual quality. It further shows generalization and robustness across diverse unseen domains.

Figures (11)

• Figure 1: Overview of motion-robust slice-to-volume MRI reconstruction. (a) Illustrates the workflow from clinical MRI acquisition to inter-slice reconstruction and downstream volumetric diagnostics. (b) Highlights key challenge of multi-type structural disruptions, including global aliasing, local detail loss, and volumetric anisotropy.
• Figure 2: Overview of strategies integrating INR variants: (a) Original INR learns a continuous image solely from sinusoidally encoded spatial coordinates. (b) Conditioned INR injects a pixel prior to guide the coordinate stream. (c) Feature-based INR integrates a feature encoder. It iteratively refines the implicit representation together with extracted image features. (d) Proposed PR-INR generalizes to 3D coordinates and performs coarse-to-continuous reconstruction via a progressive architecture. It first removes motion artifacts and global aliasing with diffusion prior, then selectively restores high-frequency details through a residual INR, and finally enforces volumetric smoothness using a continuous 3D representation. This design yields motion-robust and anatomically consistent volumetric images.
• Figure 3: Overview of the proposed PR-INR framework for motion-robust slice-to-volume MRI reconstruction. (a) Motion artifact correction. A motion-aware diffusion-based model is used to suppress intra-slice artifacts by iteratively refining image quality through forward and reverse denoising processes. (b) Local detail reconstruction. A data-consistency-aware visual encoder extracts anatomical details from under-sampled slices, guided by k-space priors and acquisition-aware feature selection. The residual INR then selectively refines high-frequency details to correct hallucinations introduced during the diffusion phase. It ensures spatial fidelity and structural consistency. (c) Inter-slice space filling. A volumetric INR takes continuous spatial coordinates and encoded features to generate dense volumetric outputs with coherent inter-slice structure.
• Figure 4: Visual comparisons of MRI reconstruction methods under different acceleration factors (AF) and motion displacements. Each subfigure shows reconstructed results from different methods. (a) Reconstruction under Max Displacement = 3 $\%$. (b) Reconstruction under Max Displacement = 5 $\%$. (c) Reconstruction with AF = 4x. (d) Reconstruction with AF = 8x. (e) Reconstruction under Max Displacement = 5 $\%$, T2-weighted, BraTS dataset. Each row includes reconstructed images (top) and error maps (bottom).
• Figure 5: Overview of the proposed multi-stage MRI reconstruction framework. (a) The motion aware diffusion process progressively removes motion artifacts and restores clean anatomical structures. (b) The implicit detail restoration process enhances local structural details and contrast in both image and k-space domains. (c) The voxel continuous-aware representation captures slice-wise continuity and reconstructs anisotropic volumes.

Theorems & Definitions (6)

• Definition 7.1: k-Space Decomposition
• Definition 7.2: PR-INR Operator
• Theorem 7.3: Variance Decomposition
• Lemma 7.4: Dual-Domain Constraint
• Theorem 7.5: Dominance of Covariance Term
• Theorem 7.6: Strict Variance Reduction