SonoNet: Despeckling of Medical Ultrasound Scans Using Convolutional Neural Network Architecture

This paper proposes a deep learning framework using convolutional neural networks (CNNs), for the despeckling in ultrasound scans. The SonoNet model proposed in this paper consists of seven layers deep net along with activation function and batch normalization. Training of the network is carried out by utilizing synthetically modeled ultrasound frames with varying speckle content. The network uses the assumption of multiplicative noise to adopt a division approach for filtering speckle content from input frames. Finally, edge features are extracted using the Canny operator and enhanced using a twodimensional 3×3 convolutional mask. Thus, the proposed CNN method effectively accomplishes speckle elimination and visual feature preservation without inducing much blurring effects. The subjective evaluation by four subject matter experts and experimental analysis using three image quality metrics indicate that the proposed system gives a superior performance as compared to the state-of-the-art methods in the field, and therefore could find applications designed for speckle suppression in ultrasound imaging systems and real-time ultrasound video segmentation and classification algorithms.