VQ-MRISR: Vector Quantization-guided referencebased MRI Super-Resolution with Deformable-Differential Cross Attention

Accruing high-resolution (HR) Magnetic Resonance Imaging (MRI) scans remains a critical challenge in clinical practice, limited by long scan times, patient motion, and hardware limitations. Super-resolution (SR) techniques have emerged as promising solutions, particularly multicontrast MRI SR (MCMSR) methods, which improve MRI scan resolution over single-contrast SR (SCMSR) by using an additional contrast image as a reference. However, the intensity inconsistency between reference and target images introduces significant challenges in MCMSR, including feature misalignment, inconsistent textures, and model confusion. In this paper, we propose VQ-MRISR, a novel reference-based MRI SR framework that integrates vector quantization (VQ) within a deep learning architecture to generate highly accurate reference images with the same contrast of the input LR inputs. To effectively exploit the reference images features, we introduce a Deformable-Differential Cross-Attention (D2CA) module that adaptively aligns both the reference and the target features for an effective information transfer. Furthermore, we employed a High- Reference Details Extraction (HRDE) module and Hierarchical Residual Channel Attention (HRCA) mechanisms to enhance both structural detail and contextual texture reconstruction. Comprehensive experiments demonstrate the superiority of VQ-MRISR against SCMSR and MCMSR methods in both quantitative metrics and visual quality. Furthermore, the use of VQ significantly reduces reference image storage demands, offering a practical solution for clinical MRI SR applications.