Diffusion Models for High-Resolution Medical Image Reconstruction and Denoising using Artifical Intellegence

Abstract

Diffusion probabilistic models (DPMs) have recently emerged as a transformative framework in image generation and restoration tasks, outperforming traditional approaches across several metrics. Their inherent capacity to iteratively denoise samples from Gaussian noise aligns closely with the inverse problems found in medical imaging, making them highly suitable for reconstruction and denoising applications. This study investigates the performance of a novel DPM-based pipeline, Diffusion-UNet, across three medical imaging modalities low-dose CT, undersampled MRI, and point-of-care ultrasound using two multi-institutional datasets totaling 4,200 volumetric images. Quantitative comparisons were made against conventional iterative reconstruction, Pix2Pix GAN, and transformer-based Restormer models. Diffusion-UNet achieved statistically significant improvements across all key image quality metrics: PSNR (42.5 dB vs. 38.1 dB, p < .001), SSIM (0.931), and NMSE (2.1 × 10⁻³). Moreover, the model demonstrated a 69.3% reduction in Fréchet Inception Distance (FID-Med), indicating enhanced perceptual realism. A blinded radiologist panel scored Diffusion-UNet reconstructions highest (κ = .84), citing better preservation of vascular structures and pathology-critical features. An ablation study on diffusion steps revealed that performance gains plateau beyond 800 steps, informing practical deployment configurations. While inference time is higher than CNNs (120 ms vs. 65 ms per slice), it remains within clinical tolerances for post-processing applications. The findings substantiate DPMs as not only technically superior but clinically viable solutions for high-resolution medical image restoration, paving the way for safer, faster, and more accurate diagnostic imaging workflows.