Systematic Review and Meta-analysis of AI-driven MRI Motion Artifact Detection and Correction

TL;DR

This systematic review and meta-analysis tackles the challenge of AI-driven MRI motion artifact detection and correction, focusing on deep learning and generative approaches. By aggregating 71 studies, it documents trends in datasets, architectures, and evaluation metrics, highlighting the prominence of GANs and diffusion models along with supervised and unsupervised training paradigms. Key findings show substantial promise for improving image quality and reducing repeat scans, but also emphasize critical hurdles: generalizability across sequences, reliance on paired data, and the risk of hallucinations, necessitating standardized datasets and reporting protocols. The work advocates for robust data sharing, zero-shot and transfer learning, and standardized benchmarks to accelerate clinical adoption and maximize the impact on diagnostic accuracy and healthcare efficiency.

Abstract

Background: To systematically review and perform a meta-analysis of artificial intelligence (AI)-driven methods for detecting and correcting magnetic resonance imaging (MRI) motion artifacts, assessing current developments, effectiveness, challenges, and future research directions. Methods: A comprehensive systematic review and meta-analysis were conducted, focusing on deep learning (DL) approaches, particularly generative models, for the detection and correction of MRI motion artifacts. Quantitative data were extracted regarding utilized datasets, DL architectures, and performance metrics. Results: DL, particularly generative models, show promise for reducing motion artifacts and improving image quality; however, limited generalizability, reliance on paired training data, and risk of visual distortions remain key challenges that motivate standardized datasets and reporting. Conclusions: AI-driven methods, particularly DL generative models, show significant potential for improving MRI image quality by effectively addressing motion artifacts. However, critical challenges must be addressed, including the need for comprehensive public datasets, standardized reporting protocols for artifact levels, and more advanced, adaptable DL techniques to reduce reliance on extensive paired datasets. Addressing these aspects could substantially enhance MRI diagnostic accuracy, reduce healthcare costs, and improve patient care outcomes.

Figures (7)

• Figure 1: Trajectory plot illustrating the evolution of samples from a three-mode Gaussian mixture distribution as they transition to a standard normal distribution $\mathcal{N}(\mathbf{0}, \mathbf{I})$. Here we simulated the Gaussian mixture as $\sum_{i=1}^{3} \omega_i \mathcal{N}(\mu_i, \sigma_i)$ with $\mu = [-2, 2, 4]^T$ and $\sigma = [0.5, 0.5, 0.5]^T$ and forward diffusion step was performed over $T=1000$ steps.
• Figure 2: Motion simulation methods are illustrated. (A) The abrupt brain movement involves rigid motion and rotation, simulated by randomly sampling translation $T_{x,y,z}$ and rotation $R_{\theta_i}$. (B) The abdominal and cardiovascular coherent movement simulation introduces blurring to the images. Motion simulation methods for static and dynamic MR images are shown. The center of k-space data, shown by red slabs, is excluded from modifications.
• Figure 3: Illustration of the overall DL-based MoCo and MoDe models. (A) left depicts an image-based training method that uses motion-corrupted and motion-free images, with an alternative approach focusing on reconstructing residual errors to generate motion-free images (red dashed line). (A) right shows an enhancement to this method using auxiliary images (data) to enhance the previous method. (B) presents motion estimation-based models, where a DNN in the left panel estimates the rigid motion parameters and another DNN in the right panel reconstructs the deformation vector field (DVF). This approach also includes the application of an DNN for DVF reconstruction from under-sampled k-space, crucial for real-time target tracking in image-guided radiation therapy.(C) illustrates model-based methods trained on k-space data. The left panel uses an DNN to estimate the amount of rigid motion, which is then used to reconstruct motion-free images iteratively. The right panel illustrates a method that unrolls the model into two modules, a denoiser DNN and data consistency module for motion artifact correction. The estimated motion parameters are used to correct the motion artifacts. (D) illustrates other types of motion-correction methods, including multi-task learning and quantitative MRI. (E) represent motion detection method, which are used either to detect motion artifacts or to select appropriate downstream tasks, such as motion-robust image reconstruction.
• Figure 4: Information of the reviewed studies. (A) The preferred reporting items for systematic reviews and meta-analyses (PRISMA) flowchart of this meta-analysis study liberati2009prisma. (B) Number of publications by research institute across years, illustrating the major contributing institutions and highlighting leading research hubs in DL-based MoCo and MoDe. (C) Number of publications by country, showing the global distribution and growth of research activity, which provides context on regional contributions and opportunities for collaboration. (D) Number of studies included for each type of statistical analysis performed in this review.
• Figure 5: Dataset information from the reviewed studies. (A) The total number of publications using institutional, public, and both types of datasets over time is illustrated. The Pearson correlation between the number of publication uses public, private, and both with time are $\rho ^p$, $\rho ^i$, and $\rho ^b$, respectively. Abbreviation: NS: Not specified. (B) and (C) Training and testing sample size are illustrated over time. The Kruskal-Wallis p-values were reported for each box plot. (D) MRI sequences used by the studies grouped by year of publication are illustrated. Abbreviations: FLAIR: fluid-attenuated inversion recovery, T1c: postcontrast T1-weighted, PD: proton density, DWI: diffusion-weighted imaging. (E) Anatomical region of the images used by the studies. The arrow shows the usage trends over time and the numbers inside parenthesis represents the slope value obtained from linear regression represents the rate of change in the number of publications per unit change in the year.