Image enhancement for accelerated MRI using a joint GAN and diffusion model framework

BACKGROUND: Magnetic resonance imaging (MRI) is a defining feature of magnetic resonance image-guided radiotherapy (MRigRT), providing superior soft-tissue contrast compared to conventional cone beam CT (CBCT). However, conventional MRI protocols typically require prolonged acquisition times. High-speed scanning techniques implemented to improve clinical workflow efficiency often produce low-quality images, characterized by reduced resolution and blurring, which compromise the accuracy of downstream tasks. Although deep learning-based enhancement methods have shown promise, most fail to adequately correct residual errors between predicted and actual images, introducing uncertainties in clinical decisions. PURPOSE: This study aims to propose and evaluate a deep learning-based approach to enhance the image quality of accelerated MRI acquisitions, with the dual objectives of ensuring registration accuracy for precise tumor targeting, and shortening in-room time to minimize patient discomfort on the treatment couch. METHODS: We acquired 72 paired 3D T2-weighted MRI scans from 62 glioma patients using a 1.5T Unity MR-Linac, with each pair comprising a standard protocol scan and an accelerated protocol scan. We proposed Residual Refinement Enhancement Network (RRENet), a novel end-to-end deep learning framework that combines generative adversarial network (GAN) and diffusion model (DM) for accelerated MRI enhancement. The framework operates in a two-stage process: first, a GAN-based predictor generates the low-frequency content; subsequently, a DM-based module refines the output by estimating residual errors. To further improve performance, an additional High-frequency Separation Training Module (HSTM) was incorporated to preserve fine anatomical details. Image quality was assessed using peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and root mean square error (RMSE). Clinical applicability was evaluated through rigid registration analysis with commercial software. RESULTS: rotation) to planning CT, ensuring the accuracy of patient setup correction. CONCLUSIONS: The proposed method can generate high-quality images from high-speed scanning sequences, reducing acquisition time while ensuring precise tumor targeting through reliable image registration.