Suppressing Speckle Noise in Ultrasound Images Through Multi-scale Structural Information Distillation

When ultrasound waves propagate through the human body, they produce reflections and scatterings, resulting in speckle noise in ultrasound images. Additionally, these speckles, being multiplicative noise, present a significant challenge for despeckling processing. Existing methods can suppress speckle noise, but they often struggle to adequately protect details while smoothing the images. In this paper, we propose a novel algorithm for ultrasound images by using multi-scale structural information distillation. Firstly, a denoising network is utilized to efficiently reduce speckle noise in ultrasound images. Next, a pre-trained network is leveraged to extract multi-scale features from both the noise ultrasound image and denoised ultrasound image. Multi-scale feature extraction is crucial as it enables the network to grasp diverse structural details across various resolutions, facilitating a deeper comprehension of the image content. Lastly, based on these extracted multi-scale features, a structural distribution similarity metric loss function is constructed. This loss function serves as a critical guide during the optimization process of the speckle noise suppression network. By minimizing this loss function, the network is trained to balance noise suppression with the preservation of fine structural details, aiming to overcoming the challenges posed by multiplicative speckle noise in ultrasound images. The proposed method is evaluated on a open access ultrasound image dataset, and the experimental results demonstrate the effectiveness of the proposed method.