A Transducer-Adaptive Denoising Model for Medical Ultrasound Imaging

Ultrasound imaging is a pivotal diagnostic tool in medical practice due to its noninvasive nature, low cost, and real-time imaging capabilities. However, the images are often affected by various types of noise, significantly degrading image quality and hindering accurate diagnosis. Traditional filtering methods struggle to preserve fine image details, while machine learning and deep learning-based models often lack adaptability to different ultrasound instruments and transducer configurations. To address these challenges, we propose a novel deep learning-based ultrasound image denoising model. Our model uses a multibranch convolutional neural network (CNN) structure to configure metadata based on sensor configuration center frequency, element kerf, element width, and imaging depth adaptively adjusts the noise level and denoising intensity, effectively suppressing speckle noise and preserving texture details in images from different ultrasound instruments. It is trained and evaluated primarily on ultrasound images synthesized using coherent plane wave compounding (CPWC), chosen for its standardized and reproducible framework. Despite being trained on CPWC data, our model demonstrates competitive denoising performance for non-CPWC-based ultrasound images, attributing to the shared physical factors influencing speckle generation. The results demonstrate that our model outperforms existing denoising methods, with average SSIM, PSNR, and EI values increased by 1.67%, 1.28%, and 1.04%, respectively. Additionally, when applied to real breast ultrasound images, our denoising method achieves state-of-the-art results in downstream image classification tasks, significantly improving both accuracy and AUC values. This study holds practical significance in enhancing the quality of ultrasound images for improved clinical diagnosis.