Attentive U-Net with Physics-Informed Loss for Noise Suppression in Medical Ultrasound Images

Ultrasound (US) imaging plays a pivotal role in medical diagnostics due to its non-invasiveness and semi-real-time imaging capabilities. However, the prevalent occurrence of speckle noise artifacts, originating from RF signal interference during the US image formation, poses a significant challenge by distorting crucial anatomical details. To address this problem, we propose an optimized U-Net model integrated with gated attention and trained using a physics-informed loss function. This approach enables the network to adeptly suppress noise distortions in US images during training. The integration of the attention mechanism is to facilitate effective capturing of long-range feature representations, while the physics-informed loss function is implemented to promote rapid convergence and self-adaptive learning, aligning with the physical constraints of interpretable US images with clinical relevance. Experimental validation conducted on the US-4 and BUSI datasets demonstrates competitive performance, indicating significant reduction of speckle noise. This study presents a niche methodology in the domain of medical image denoising aimed at improving US diagnostic accuracy and clinical outcomes through AI-driven advancements.