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

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Magnetic resonance imaging (MRI) offers superior soft-tissue contrast in image-guided radiotherapy (MRgRT) compared to conventional CT.
• Accelerated MRI protocols improve workflow but often yield low-quality images, impacting accuracy.
• Existing deep learning methods struggle to fully correct errors, leading to clinical uncertainties.

Purpose of the Study:

• To develop and assess a deep learning approach for enhancing accelerated MRI image quality.
• To ensure accurate tumor targeting through precise image registration.
• To reduce patient scan time and discomfort during radiotherapy.

Main Methods:

• A novel deep learning framework, Residual Refinement Enhancement Network (RRENet), was developed, combining generative adversarial networks (GANs) and diffusion models (DMs).
• RRENet employs a two-stage process: GAN for low-frequency content and DM for residual error refinement.
• A High-frequency Separation Training Module (HSTM) was integrated to preserve fine anatomical details, with image quality assessed via PSNR, SSIM, and RMSE.

Main Results:

• RRENet reduced imaging time by 70% compared to standard protocols while maintaining high image quality (SSIM: 92.27%, PSNR: 32.84dB).
• Enhanced images demonstrated accurate alignment with planning CT (≤2.2 mm translation, ≤0.2° rotation).
• The method ensured accuracy for patient setup correction in MRgRT.

Conclusions:

• The proposed RRENet method generates high-quality images from accelerated MRI sequences.
• It effectively reduces acquisition time, crucial for MRgRT workflows.
• The approach ensures precise tumor targeting via reliable image registration.