In-silico comparison of a diffusion model with conventionally trained deep networks for translating 64mT to 3T brain FLAIR

Deep learning (DL) methods are increasingly applied to address the low signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) of low-field MRI (LFMRI). This study evaluates the potential of diffusion models for LFMRI enhancement, comparing the Super-resolution via Repeated Refinement (SR3), a generative diffusion model, to traditional architectures such as CycleGAN and UNet for translating LFMRI to high-field MRI (HFMRI). Using synthetic LFMRI (64mT) FLAIR brain images generated from the BraTS 2019 dataset (3T), the models were assessed with traditional metrics, including structural similarity index (SSIM) and normalized root-mean-squared error (nRMSE), alongside specialized structural error measurements such as gradient entropy (gEn), gradient error (GE), and perception-based image quality evaluator (PIQE). SR3 significantly outperformed (p-value < < 0.05) the other models across all metrics, achieving SSIM scores over 0.97 and excelling in preserving pathological structures such as necrotic core and edema, with lower gEn and GE values. These findings suggest diffusion models are a robust alternative to conventional DL approaches for LF-to-HF MRI translation. By preserving structural details and enhancing image quality, SR3 could improve the clinical utility of LFMRI systems, making high-quality MRI more accessible. This work demonstrates the potential of diffusion models in advancing medical image enhancement and translation.