Real-Time Ultrasound Image Despeckling Using Mixed-Attention Mechanism Based Residual UNet
Yancheng Lan, Xuming Zhang
Abstract
Ultrasound imaging has been widely used for clinical diagnosis. However, the inherent speckle noise will degrade the quality of ultrasound images. Existing despeckling methods cannot deliver sufficient speckle reduction and preserve image details well at high noise corruption and they cannot realize real-time ultrasound image denoising. With the popularity of deep learning, supervised learning for image denoising has recently attracted considerable attention. In this paper, we have proposed a novel residual UNet using mixed-attention mechanism (MARU) for real-time ultrasound image despeckling. In view of the signal-dependent characteristics of speckle noise, we have designed an encoder-decoder network to reconstruct the despeckled image by extracting features from the noisy image. Furthermore, a lightweight mixed-attention block is proposed to effectively enhance the image features and suppress some speckle noise during the encoding phase by using separation and re-fusion strategy for channel and spatial attention. Besides, we have graded the speckle noise levels with a certain interval and designed an algorithm to estimate the noise levels for despeckling real ultrasound images. Experiments have been done on the natural images, the synthetic image, the image simulated using Field II and the real ultrasound images. Compared with existing despeckling methods, the proposed network has achieved the state-of-the-art despeckling performance in terms of subjective human vision and such quantitative indexes as peak signal to noise ratio (PSNR), structural similarity (SSIM), equivalent number of looks (ENL) and contrast-to-noise ratio (CNR).