Removal of Speckle Noises from Ultrasound Images Using Parallel Convolutional Neural Network

Speckle noises widely exist in ultrasound images. They seriously affect the quality of images and cause the doctor to make mistakes in diagnosis. In this paper, we propose a three-path parallel convolutional neural network called USNet to achieve speckle reduction for ultrasound images. We combine three different sub-networks to increase the width of the whole network instead of the depth. The ideas of dilated convolution and shortcut connection are added to increasing the learning ability of the model. These make our proposed USNet can learn deeper information from the original ultrasound image. In experiments, we verify the effectiveness of the proposed three-path parallel structure and the dilated convolution by conducting ablation experiments. At the same time, we propose a different method to construct the training dataset. For the noisy training image, we artificially add speckle noise at three different σ levels to enhance the generalization performance of the proposed method. For the noise-free true labels, we use two stages to obtain on the basis of original images, including Optimized Bayesian Non-Local Means with block selection method (OBNLM) and Second-order Oriented Partial-differential Equation (SOOPDE) method. Then we compare our proposed method with other four different methods, including Kuan method, Speckle Reducing Anisotropic Diffusion (SRAD) method, the OBNLM method, and Residual Learning Network (RLNet). We qualitatively and quantitatively evaluate these methods in terms of smoothness, texture information protection, and edge clarity. The results show that our proposed USNet model can batch and quickly achieve good speckle reduction as well as texture preservation for different types of ultrasound images without any parameter adjustment. The USNet model has the advantages of good adaptability, robustness, and generalization. It is of great significance for improving the diagnostic efficiency of clinical medicine.