| In the early diagnosis of cancer diseases,the morphological characteristics of cell nuclei in pathological images are one of the most important discriminative criteria and are often used to predict the treatment effect.Therefore,the accurate segmentation of cell nuclei in pathological images becomes the basis of medical image analysis and processing.Although U-Net network is widely used in the field of medical image segmentation,it still faces problems such as inability to obtain fine features and easy loss of details,which lead to less than ideal segmentation results.For this reason,this paper carries out research on the segmentation method of pathological images based on U-Net network,and the main work is as follows:(1)An improved segmentation network,ResUNet_LPAT,is designed to address the problems of U-Net network in the detail processing of segmented images.First,a residual network is added to the U-Net network,and ResUNet is used as the improved base network to ensure that the subsequent improvements will not degrade due to the deepening of the network;second,the LeakyReLU activation function is used instead of the ReLU activation function to enable gradient learning and back propagation even for the input negative value information;then,the residual paths are composed using different size residual blocks Finally,the channel and spatial attention modules are used to suppress irrelevant information and improve the segmentation effect of the image.At the same time,two loss functions,NLLLoss and Dice Loss,are used to optimize the network together with the freeze training strategy during the training process.To verify the feasibility of the ResUNet_LPAT network,experiments are conducted on two data sets,and the results of multiple comparison experiments show that the ResUNet LPAT network improves 10.22%,14.24%,and 3.04%in the evaluation indexes of MioU,Dice,and Acc,respectively,on the A data set,and 3.01%on the B data set,respectively,compared with the U-Net network.improved by 3.01%,4.22%,and 0.7%,respectively.The evaluation metrics are also improved compared with classical segmentation networks such as SegNet and FCN8s.(2)A segmentation network ResUNet_GAN incorporating adversarial networks is designed and applied to the field of image segmentation of cell nuclei.First,borrowing the framework structure of generative adversarial network,ResUNet,which is added to the U-Net network after the residual network,is used as the generative network;second,a double-discrimination network based on Image GAN is designed to strengthen the semantic segmentation effect by adversarial training.Meanwhile,the training process uses two loss functions,NLLLoss and Dice Loss,together with the freeze training strategy to optimize the generative network.To verify the feasibility of the ResUNet_GAN network,experiments are conducted on two datasets,and the results of multiple comparison experiments show that the ResUNet_GAN network improves 9.59%,12.34%,and 2.93%in MioU,Dice,and Acc evaluation metrics on dataset A,and improves 3.07%and 3.07%on dataset B,respectively,compared with the U-Net network.by 3.07%,3.73%,and 0.69%on the B dataset,respectively.Compared with the classical segmentation networks such as SegNet and FCN8s,the evaluation indexes are also improved.The proposed ResUNet_LPAT and ResUNet_GAN segmentation networks have good generalization ability and segmentation accuracy on both data sets,which can effectively improve the accuracy of cell nuclei segmentation in pathological images and provide a basis for the diagnosis of cancer diseases. |
