SUPREM: A Super-Resolution Network for Through-Plane Structure Enhancement in Multi-Contrast MRI Using Disentangled Representation Learning

Motivation: Through-plane super-resolution in brain MRI is crucial for reducing discomfort during clinical assessments, yet current methods mainly focus on in-plane enhancements. Goal(s): This study aims to develop a novel SR method that utilizes disentangled representation learning to improve the fine details of low-resolution T2 by leveraging high-resolution T1. Approach: We introduce the SUPREM network, which reconstructs high-quality of low-resolution T2 by leveraging structural information decomposed from high-resolution T1, employing a progressive supervised reconstruction and patch-wise contrastive learning. Results: Our method demonstrates superior structural detail recovery, achieving the highest scores for SSIM and PSNR metrics across multiple datasets, indicating its effectiveness and clinical relevance. Impact: The proposed method has the potential to reduce discomfort during brain MRI assessments by enhancing through-plane super-resolution, thereby improving structural detail recovery and supporting more accurate diagnostics in clinical practice.