Denoising of Medical Ultrasound Images Using Variational Autoencoders Guided by Attention Mechanisms

Ultrasound imaging, while widely used and clinically valuable, continues to face a major limitation, speckle noise. This type of noise interferes with the visibility of fine anatomical structures and can lower the clinician's diagnostic confidence. Traditional denoising methods often struggle to strike the right balance between effectively reducing noise and preserving essential image details. As a result, their usefulness in real-world, high-demand clinical settings remains limited. To address this ongoing challenge, we introduce a new denoising framework that combines large kernel attention (LKA) with a convolutional variational autoencoder (CVAE). This hybrid approach takes advantage of LKA's ability to capture broad contextual information, improving feature representation, while the generative power of the CVAE helps reconstruct clean, speckle-free ultrasound images with greater reliability. Extensive evaluation across six diverse ultrasound datasets demonstrates that LKA-CVAE consistently outperforms state-of-the-art methods. The model achieves an average peak signal-to-noise ratio (PSNR) of 53.64 dB and structural similarity index measure (SSIM) of 0.913, representing improvements of 10 dB and 0.041, respectively, over the best competing approach. These gains are statistically significant and accompanied by minimal variability, with standard deviations of 0.028 for PSNR and 0.03 for SSIM. These results establish LKA-CVAE as a robust, high-performance tool for ultrasound image enhancement, supporting more confident and reliable clinical interpretation.