Discriminative-Constrained High-Frequency Diffusion Prior Modeling for Cerebral MRI Super-Resolution

Super-resolution (SR) reconstruction of Magnetic resonance imagin (MRI) images plays a crucial role in advancing medical imaging by enhancing the visibility of fine anatomical details. In recent years, generative models have demonstrated remarkable capabilities in SR tasks by synthesizing more realistic high-frequency information. However, generative models generally exhibit considerable randomness, making it challenging to ensure the stability and consistency of the results. To address this, we propose a novel MRI SR method that not only integrates the strengths of both generative and discriminative models, but also enables high-fidelity reconstruction of high-frequency details. Specifically, we first decouple the high- and low-frequency signals through wavelet decomposition. The high-frequency components are modeled using a latent diffusion model (LDT), while the low-frequency information from the low-resolution (LR) images is used as a conditional constraint to guide the generation of high-frequency details. Finally, the LR features and the reconstructed high-frequency features are fused by a discriminative decoder to produce the final SR image. Quantitative experimental results demonstrate that our method outperforms existing state-of-the-art MRI SR approaches across multiple metrics while maintaining a more lightweight architecture. Furthermore, visualization results further validate the superiority of our method in reconstructing high-frequency details.