EU-Net: Boundary Enhancement Based Medical Ultrasound Image Segmentation Network

U-Net segmentation network is an effective medical image segmentation model. U-Net, as a fully convolutional network, is vulnerable to the limitation that convolution cannot extract image context information well. Therefore, this paper proposes a boundary enhancement based medical ultrasound image segmentation network: EU-Net, which uses the ConvMixer module at the bottom of the encoder to extract the global contextual information of the image. And a depthwise separable convolution is used in the jump connection to help the network capture the effective features in each layer of the encoder without increasing the computational effort. Ultrasound images are a non-invasive and painless diagnostic modality, but since the ultrasound imaging process is prone to noise, the segmentation of ultrasound images is prone to unclear boundary segmentation, so an active boundary loss (ABL) is added to guide the prediction of boundary movement during each iteration, thus improving the segmentation details of the network. The evaluation on an open source breast ultrasound dataset shows that EU-Net achieves an average IOU of 0.684 and Dice of 0.809.