Multi-Contrast MRI Motion Correction via Parameter-Informed Disentanglement and Adaptive Experts

Researchers propose a unified deep learning framework for correcting motion artifacts across different MRI contrast types by combining contrast disentanglement with severity-aware adaptive correction. The method achieves measurable improvements over existing approaches and demonstrates robust generalization to unseen clinical data, addressing a key limitation where current solutions fail across diverse imaging modalities.

Motion artifacts remain a persistent challenge in magnetic resonance imaging, compromising diagnostic accuracy and forcing clinicians to repeat scans at increased cost and patient discomfort. Traditional deep learning solutions suffer from a critical vulnerability: they are trained on specific contrast types and fail when encountering different MRI modalities or artifact severities in real-world clinical settings. This research tackles the generalization problem through an architecture that separates contrast-specific information from anatomical content, enabling the same model to function effectively across diverse imaging protocols.

The technical approach leverages ScanCLIP, a pretrained vision-language model conditioned on 30,000 MRI text-image pairs, to extract contrast embeddings directly from acquisition parameters. This design choice elegantly sidesteps the need for explicit contrast labeling during inference. The routing mechanism through a Mixture-of-Experts network allows the system to adapt correction strategies based on estimated artifact severity, addressing the reality that heavy artifacts require different treatment than light ones.

For healthcare providers and medical imaging vendors, the implications are substantial. Improved artifact correction reduces the need for repeat scans, lowering operational costs and patient burden while improving diagnostic reliability. The demonstrated zero-shot generalization on unseen scanning parameters is particularly valuable, as clinical environments frequently encounter new or modified acquisition protocols. The quantitative improvements—0.75 dB PSNR and up to 0.0279 SSIM gains—may appear modest numerically but translate to meaningful diagnostic quality improvements in practice, especially at higher artifact severities where clinical impact is greatest.

• →Parameter-informed disentanglement enables MRI motion correction to generalize across diverse contrast types and acquisition protocols without contrast-specific retraining.
• →Mixture-of-Experts routing adapts correction strategies based on artifact severity, improving results particularly on heavily corrupted images.
• →Zero-shot generalization on real clinical data demonstrates robustness to unseen scanning parameters, a critical limitation of existing methods.
• →Dual-pathway decoder architecture enforces image-space consistency while reconstructing both corrected images and artifact maps.
• →Performance improvements measured on IXI and HCP benchmarks show 0.75 dB PSNR and up to 0.0279 SSIM gains over state-of-the-art approaches.