Kernel-aware network with dual diffusion model for MRI blind super resolution

Abstract Magnetic resonance imaging (MRI) offers abundant and rich information to help doctors quickly diagnose patients’ diseases. Due to patients’ movement, the obtained MRI is blurry, which may have a great influence on the accuracy of clinical diagnosis. To tackle this challenge, we propose a novel kernel-aware network with dual diffusion model (KDDM) for MRI blind super resolution, which includes two novel neural networks, namely the kernel-aware prediction network (KAPN) and the adaptive denoising reconstruction network (ADRN). KAPN is designed to accurately estimate the blur kernel by capturing its global context information. Meanwhile, the ADRN makes full use of the power learning ability of the diffusion model to reconstruct MRI images with subtle pathological features with the help of the blur kernel estimated by the KAPN. Through the collaboration between the KAPN and ADRN, the performance of our proposed KDDM is significantly enhanced. The experimental results on IXI and FastMRI datasets show that our method can produce better super-resolution images with clear edges and textures and achieve better reconstruction results than several mainstream methods.