Automatic Segmentation of Median Nerve in Ultrasound Image by a Combined Use of U-Net and VGG16

In this paper, we proposed a new method VGG16-UNet based on U-Net and VGG16 to robustly and automatically segment median nerve (MN) in ultrasound image. The proposed method has an encoder-decoder neural network architecture. The encoder is the VGG16 network removing the three fully connected layers and the decoder is the architecture resembling the right part of U-Net with changes of activation functions as well as skip connection. A dataset of 891 MN ultrasound images was acquired and used for training and evaluating the proposed method and U-Net. With three metrics including DC, JC and precision, the quantitative evaluation results showed that the proposed method outperformed U-Net in MN segmentation. The trained VGG16-UNet network was applied to segment a new MN image sequence and the 3D MN was obtained using VTK.