Edge-guided conditional diffusion model for multi-contrast MRI super-resolution

Multi-contrast MRI super-resolution assisted by auxiliary anatomical guidance has emerged as a pivotal strategy for accelerating clinical imaging protocols. While diffusion models (DMs) demonstrate promising capabilities in this domain, two critical limitations hinder their diagnostic applicability: (1) insufficient utilization of high-frequency edge features across multi-contrast inputs, leading to anatomically inconsistent boundary construction, and (2) inflexible fusion mechanisms that fail to adequately model cross-modality dependencies, resulting in structural distortions. To address these challenges, we propose an edge-guided conditional diffusion model (i.e., Eg-Diff) for multi-contrast MRI super-resolution. To preserve high-frequency anatomical details, we design high-frequency-injected encoders in which contrast-adaptive edge features, captured by learnable detection operators, are exploited for hierarchical injection. Furthermore, to promote multi-modality synergy, adaptive multi-modality feature fusion is developed. Through contrast-weighted gating and K-selective balanced attention, it dynamically calibrates auxiliary guidance, enabling precise auxiliary prior utilization while suppressing modality bias. In the reverse diffusion process, to minimize noise interference, we propose a proactive activation strategy that discriminatively integrates the edge features as prior conditions for diffusion. Extensive experiments on IXI and MRBrainS13 datasets demonstrate that Eg-Diff outperforms state-of-the-art methods, highlighting its potential for clinical applications.