Abstract:
Ultrasound imaging is widely used in clinical practice due to its real-time visualization, safety, and cost-effectiveness; however, its diagnostic quality is often hindered by speckle noise and acoustic clutter, which obscure anatomical boundaries and reduce interpretability. Traditional denoising methods—such as Gaussian, Median, and Non-Local Means filtering—struggle to address the highly non-linear and spatially variant noise characteristics inherent to ultrasound data. To overcome these limitations, this research introduces UltraDiffusion, a novel denoising framework built upon Denoising Diffusion Probabilistic Models (DDPMs), a class of generative models that has demonstrated state-of-the-art performance in medical image synthesis and reconstruction. UltraDiffusion leverages selective fine-tuning of the UNet backbone in Stable Diffusion v1.5, originally trained on over eight billion natural images. The model was adapted using a normalised and speckle-augmented subset of the Kaggle Ultrasound Nerve Segmentation dataset, with additional training optimizations including mixed-precision computation, component freezing, and deployment on T4 and RTX 4090 GPUs.
Quantitative evaluation using PSNR, SSIM, and LPIPS shows that UltraDiffusion consistently outperforms conventional filtering techniques, delivering superior noise suppression while preserving structural details critical for clinical assessment. Notably, the model achieved robust performance despite being trained on only 5,000 samples, demonstrating its capacity to generalize and effectively reconstruct fundamental anatomical patterns. These results underscore the potential of DDPM-based generative modelling as a powerful tool for ultrasound enhancement. UltraDiffusion provides a promising foundation for future research in diagnostic quality improvement, automated image interpretation, and broader clinical applications of diffusion models in medical imaging.
