A U-Net Network Model for Medical Image Segmentation Based on Improved Skip Connections

Abstract

To address the loss problem introduced by downsampling in the classical U-Net architecture, this paper improves the U-Net model and uses the model for medical image segmentation. The essence of the proposed model is still the classical U-Net encoder-decoder network, in which the encoder and decoder sub-networks are connected by a skip connection. First, we improve the connection position of the skip connection, and the two ends of the connection are changed from the second convolution result of the original convolution block to the first result and the decoder convolution block for concatenation; second, the concatenation operation is added to the convolution block of the downsampling part, and the two improvements aim at retaining more image underlying information, and thus achieving more efficient fusion of high and low level image information; finally, on the public medical image segmentation dataset, the classical U-Net, FCN-8s and the improved model in this paper are comparatively evaluated for cell nucleus segmentation in microscope images and liver segmentation in abdominal CT scans. The experiments show that the improved U-Net model mIoU, Aver_dice in this paper improves by 2~3% compared with the control model.