Ultrasound vascular image segmentation based on full convolutional aggregation neural network

Abstract

The concept of ultrasound therapy has been proposed long ago. However, most of the previous methods of ultrasound therapy destroy tissue through thermal effects and cause great damage to patients. In recent years, ultrasound cavitation therapy has caused extensive discussion and research due to its unique non-invasiveness. In order to achieve real-time access to the patient during treatment, the lesion area must be processed simultaneously and accurately. However, ultrasound images mostly have lower signal-to-noise ratio, contrast, and blurred edges. In order to solve the segmentation problem, this paper proposes a new segmentation model Du-net based on the full convolutional neural network. Under the premise of deepening the network depth to obtain more information, the Encoder-Decoder method and the layered aggregation mode are used to prevent the gradient explosion. Solving boundary segmentation problem by constructing joint network, In the case of a small data set, the data enhancement method using a suitable data set effectively increases the usable image features and achieves better results on the ultrasound blood vessel image.