SE-UXNet: A Medical Imaging Segmentation Network Based on Attention Mechanism

Abstract

Automatic medical image segmentation is an important part of medical image analysis, and plays an indispensable role in computer-aided diagnosis. Recently, FCN (Fully Convolutional Network) and U-Net have become the main frameworks for image segmentation task. Based on U-Net, the U-Net++ is proposed to predict robust segmentation maps. However, it is difficult for both U-Net and U-Net++ to obtain better results as the model depth grows. In this paper we propose a new medical image segmentation network SE-UXNet using attention mechanism. This model bridges the encoder-decoder models of two different structures to increase model capacity, and recalibrates the weight of the features in the bridge process by SE blocks. Using the Dice coefficient and mIoU as evaluation criteria, we evaluate our model by performing nuclei, liver, prostate, polyps and pulmonary nodule segmentation tasks. The results show that, the Dice coefficients of the above segmentation tasks using SE-UXNet were 92.75%, 90.49%, 90.34%, 95.33% and 73.44%, respectively, all higher compared to results of the U-Net and U-Net++. The experimental results also show that the method we proposed improves the accuracy of medical image segmentation.