Diffusion-Based Enhancement of Low-Quality MRI Images for Improved Diagnostic Accuracy

Medical imaging modalities such as MRI and CT often suffer from low signal-to-noise, low contrast, and limited spatial resolution due to hardware constraints and acquisition time. This paper proposes a diffusion-based generative model to enhance low-quality medical images. Unlike traditional CNN or GAN approaches, we leverage a denoising diffusion probabilistic model (DDPM) variant that learns to iteratively refine noisy, low-resolution inputs into high-quality outputs. Our model is trained on public datasets (e.g. BraTS brain MRIs and NIH Chest X-rays) with simulated degradation (added noise and downsampling). We introduce a medically-inspired loss (combining reconstruction loss with structural similarity) and use accelerated sampling (e.g. DDIM) to reduce inference time. In experiments, the diffusion model substantially outperforms strong baselines: it achieves higher PSNR/SSIM and visibly sharper detail than a CNN super-resolution network and an SRGAN. For example, on brain MRI slices our method attains PSNR ≈38.8 dB and SSIM ≈0.95 versus 29.5 dB/0.88 for bicubic and 32.5 dB/0.90 for a GAN-based model. These improvements translate to clearer anatomical features, which can improve diagnostic accuracy. We discuss the implications for accelerating clinical workflows and outline future work (e.g. faster sampling and rigorous clinical validation).