Research on Optimization of cell Nucleus Image Segmentation based on U-Net

Abstract

To solve the problem of low precision of nucleus segmentation in cell microscopic images, a method of nucleus segmentation based on deep learning is proposed. This paper presents a new network architecture Res18 U-Net++, which is based on the network framework of U-Net++ and introduces the residual module ResNet-18 and spatial attention mechanism. The network structure can make better use of the relationship between feature maps, and use set modules to enhance the reuse of feature information, thus improving the network performance. This work uses a large number of training data sets to train the network. Numerical calculation and experimental results show that this method can quickly and accurately achieve the segmentation of cell nucleus images, and achieve the effect of data enhancement. On this basis, the parameters of the model such as optimizer, loss function, learning rate, iteration value, and batch_size are modified and debugged, and images are preprocessed to try to increase image pixels to achieve better training results. In addition, the network framework proposed in this paper is compared with many network structures in the U-Net family, and the higher IoU value and smaller loss function show the advantages of the network framework proposed in this paper. Finally, validation nuclei segmentation is performed using the generative adversarial network trained on the validation dataset, demonstrating the feasibility of the method. The U-Net cell nucleus image segmentation method based on the convolutional neural network proposed in this paper is helpful for cell biology research and medical image cell nucleus processing.