A High-Frequency Re-Optimization Network for MRI Reconstruction with CT as the Prior

Prior-constrained deep learning (DL)-based reconstruction can reconstruct MR images with high quality. However, such methods face the inherent risk of introducing hallucinations to the reconstructed images because of the direct concatenation of the prior. Furthermore, most of these methods solely leverage MR images as the prior information, often overlooking the potential advantages of integrating data from different imaging modalities characterized by high resolution and fast imaging speed, such as Computed Tomography (CT). In this study, we introduced a robust High-Frequency Re-optimization Network (HFReopt-Net) to iteratively reconstruct MR images with CT as the prior. Leveraging a sparsifying transform commonly employed in MRI reconstruction and guided by the intrinsic characteristics of the target image, our proposed network can decompose the input MR image into low-frequency and high-frequency components. While the low-frequency components are selectively preserved, the high-frequency components that contain the details of the target image can be recovered by the network through the prior information of CT. Experimental results demonstrated the robustness of the proposed method and its ability to achieve reliably reconstructed MR images with high quality.