Wavelet-Optimized Motion Artifact Correction in 3D MRI Using Pre-trained 2D Score Priors

TL;DR

This work tackles MRI motion artifacts, a major diagnostic hindrance, by moving beyond traditional 3D diffusion models that rely on explicit forward degradations. It introduces 3D-WMoCo, which combines two orthogonal pre-trained 2D score priors, a mean-reverting stochastic differential equation for restoration, and wavelet-based diffusion and convolutions to accelerate inference. The approach yields superior artifact removal and structural preservation on both simulated and real 3D MRI data, with notable gains in inter-slice coherence and processing speed. The method generalizes to other 3D restoration tasks and offers a practical, scalable framework for clinical translation without requiring explicit degradation modeling.

Abstract

Motion artifacts in magnetic resonance imaging (MRI) remain a major challenge, as they degrade image quality and compromise diagnostic reliability. Score-based generative models (SGMs) have recently shown promise for artifact removal. However, existing 3D SGM-based approaches are limited in two key aspects: (1) their strong dependence on known forward operators makes them ineffective for correcting MRI motion artifacts, and (2) their slow inference speed hinders clinical translation. To overcome these challenges, we propose a wavelet-optimized end-to-end framework for 3D MRI motion correct using pre-trained 2D score priors (3D-WMoCo). Specifically, two orthogonal 2D score priors are leveraged to guide the 3D distribution prior, while a mean-reverting stochastic differential equation (SDE) is employed to model the restoration process of motion-corrupted 3D volumes to motion-free 3D distribution. Furthermore, wavelet diffusion is introduced to accelerate inference, and wavelet convolution is applied to enhance feature extraction. We validate the effectiveness of our approach through both simulated motion artifact experiments and real-world clinical motion artifact correction tests. The proposed method achieves robust performance improvements over existing techniques. Implementation details and source code are available at: https://github.com/ZG-yuan/3D-WMoCo.

Figures (10)

• Figure 1: Results of CNN-based and SGM-based motion correction methods trained on the sagittal plane. Yellow arrows highlight residual motion artifacts, while green arrows indicate inter-slice inconsistencies.
• Figure 2: Overview of the Proposed 3D-WMoCo. The diffusion endpoint of the proposed model is a motion-corrupted volume ${\bf{\mu }}$ with added Gaussian noise ${\cal N}(0,{\lambda ^2})$. We perform the restoration process from a motion-corrupted MRI volume ${\bf{v}}(T)$ to a motion-free MRI volume ${\bf{v}}(0)$ by alternately using the scores obtained from two orthogonal slices ${s_{{\phi ^ * }}}$ and ${s_{{\varphi ^ * }}}$. The model is accelerated and optimized by incorporating wavelet transforms.
• Figure 3: The model is implemented in the wavelet domain. The image is decomposed into four wavelet components, which are then concatenated along their dimensions.
• Figure 4: The basic structure of the wavelet residual block is illustrated. Conv denotes the basic convolution block, WTConv is the wavelet convolution block, and SiLU is the activation function.
• Figure 5: Qualitative results of mild motion artifact correction for brain data in sagittal, coronal, and horizontal plane views using various methods. Yellow and green ROIs are presented as local magnifications. The first column denotes the ground truth (GT), the second column denotes the low-quality image (LQ) corrupted by motion artifacts, the third to sixth columns denote results of comparison algorithms, and the last column denotes our proposed method.