A More Streamlined U-net for Nerve Segmentation in Ultrasound Images
Yun Wang, Chang Wei, Zhijian Wang, Qinggao Lu, Chenggang Wang · 2018
U-net [1] is used in medical image segmentation [2]. Its core ideal is to build a concatenation [3] between the feature map of upsampling from the expansive path and the correspondingly cropped feature map from the contraction path. The cropping is very important because of the loss of border pixels when convolution is performed. However, we think that too many concatenations will bring too many cropping and convolution operations, which will reduce the information of the image, and will ultimately weaken the segmentation effect. So we remove a concatenation between the expansive path and the contracting path, and a new network called mini-u-net is proposed. In order to prove the superiority of our network, we have also created a network with one-more concatenation, we call it onemore-u-net. We put onemore-u-net, original u-net, and mini-u-net for comparison. The dataset uses the neural image data of the kaggle2016 game. The experimental comparison found that the dice-coef', precision, recall of mini-u-net are all better than the other two networks, and then we put The mini-u-net compares the results of two papers which use the same dataset and finds that our results in precision and dice-coef are better than those mentioned in the paper.