An Improved U-Net Based on Self-adaptive Difference Convolution

Medical image segmentation plays an important role in medical applications, such as clinicians' diagnosis, treatment and disease evaluation. In this paper, a self-adaptive difference U-Net based on the variable proportion of local gradient and intensity information is proposed to segment medical images. The proposed approach aggregates intensity information with gradient information dynamically by the attention mechanism. The dynamic proportion factors of gradient information are assigned to each channel so that the network performs well in the segmentation of medical images. Experimental results demonstrate that the proposed method outperforms other U-Net based segmentation networks with respect to the accuracy, AUC, precision and specificity, etc.