IM-MoCo: Self-supervised MRI Motion Correction using Motion-Guided Implicit Neural Representations

TL;DR

IM-MoCo addresses MRI motion artifacts by combining motion-detection with a pair of motion-guided implicit neural representations. The method jointly optimizes an Image INR and a Motion INR under a data-consistency objective, guided by a k-space line mask predicted by kLD-Net, and regularized by a gradient-entropy prior; the forward model for motion is $K = \sum_{t=1}^T S_t \odot \mathcal{F} \mathbf{C} \mathbf{M}_t I$, enabling end-to-end reconstruction that accounts for 2D rigid motion. On NYU fastMRI data with simulated motion, IM-MoCo outperforms AF, U-Net, and AF+ across SSIM, PSNR, and HaarPSI (e.g., Light motion: SSIM up to $98.25\pm1.25$, PSNR $40.06\pm3.33$ dB, HaarPSI $97.20\pm4.05$; Heavy motion: SSIM $92.77\pm3.59$, PSNR $33.06\pm3.59$ dB, HaarPSI $87.29\pm9.38$). It also improves downstream pathology classification accuracy, achieving $97.06\%$ (light) and $96.32\%$ (heavy) on fastMRI+ patches. These results suggest motion-guided INRs can robustly suppress motion artifacts while preserving clinically relevant structures, with potential for extension to multi-coil data and real-motion scenarios.

Abstract

Motion artifacts in Magnetic Resonance Imaging (MRI) arise due to relatively long acquisition times and can compromise the clinical utility of acquired images. Traditional motion correction methods often fail to address severe motion, leading to distorted and unreliable results. Deep Learning (DL) alleviated such pitfalls through generalization with the cost of vanishing structures and hallucinations, making it challenging to apply in the medical field where hallucinated structures can tremendously impact the diagnostic outcome. In this work, we present an instance-wise motion correction pipeline that leverages motion-guided Implicit Neural Representations (INRs) to mitigate the impact of motion artifacts while retaining anatomical structure. Our method is evaluated using the NYU fastMRI dataset with different degrees of simulated motion severity. For the correction alone, we can improve over state-of-the-art image reconstruction methods by $+5\%$ SSIM, $+5\:db$ PSNR, and $+14\%$ HaarPSI. Clinical relevance is demonstrated by a subsequent experiment, where our method improves classification outcomes by at least $+1.5$ accuracy percentage points compared to motion-corrupted images.

Figures (6)

• Figure 1: IM-MoCo pipeline. The pre-trained $k$lD-Net takes a motion-corrupted $k$-space and outputs a motion mask, which is post-processed to yield a list of movement groups indicated by the different colors in $S$ for better visualization. The number of movements and corresponding lines guide the Motion INR and the Image INR. The Image INR predicts the motion-free image and the Motion INR is used for guidance by optimizing the forward motion yielding a motion-corrupted $k$-space. The discrepancy between the motion-corrupted and the measured $k$-space is minimized using the data-consistency (DC) loss. The gradient entropy is a denoiser on the Image-INR output to impose crisp image priors. The final motion-corrected image is the output of the Image INR.
• Figure 2: The visualization shows the median results of motion-corrected images of our IM-MoCo pipeline besides motion-corrupted, ground truth, and comparison methods. The first and third rows show the light and heavy correction results, respectively. The second and fourth rows show the residual error images.
• Figure 3: $k$lD-Net detection results. The first and second show the detection performance of the $k$LD-Net for light and heavy motion, respectively.
• Figure 4: The visualization shows the worst results of motion-corrected images of our IM-MoCo pipeline besides motion-corrupted, ground truth, and comparison methods. The first and third rows show the light and heavy correction results, respectively. The second and fourth rows show the residual error images.
• Figure 5: The visualization shows the best results of motion-corrected images of our IM-MoCo pipeline besides motion-corrupted, ground truth, and comparison methods. The first and third rows show the light and heavy correction results, respectively. The second and fourth rows show the residual error images.