Title: Adaptive Expert Mechanisms and Parameter-Guided Disentanglement for Multi-Contrast MRI Motion Correction

Abstract: Diagnostic accuracy in magnetic resonance imaging (MRI) is frequently compromised by motion artifacts. Current deep learning solutions are generally limited to specific image contrasts, lacking the ability to generalize across varying modalities and degrees of artifact severity. To address these limitations, we introduce a comprehensive framework that integrates severity-adaptive correction with parameter-driven contrast disentanglement. By leveraging ScanCLIP, a model pretrained on more than 30,000 text-image pairs of MRIs, our approach extracts contrast embeddings directly from acquisition parameters. This process separates contrast styling from anatomical structures, producing contrast-agnostic features. Subsequently, a Vision Transformer assesses the level of motion and directs these features through a Mixture-of-Experts architecture to facilitate precise artifact removal. The system employs a dual-pathway decoder to reconstruct both the corrected image and a residual artifact map, thereby ensuring consistency within the image space. Evaluations on the IXI and HCP datasets show that our method surpasses current state-of-the-art techniques, achieving a PSNR increase of 0.75 dB and an SSIM improvement of up to 0.0279, with performance gains becoming more pronounced as artifact severity increases. Furthermore, the model exhibits strong zero-shot generalization capabilities on real-world clinical data acquired using previously unseen scanning parameters, whereas existing methods typically either fail to eliminate artifacts or generate new distortions.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC