Research On Cell Image Segmentation Algorithm Based On CycleGAN Network

The segmentation of cervical cell image is an important part in the detection of cancer cells.At present,traditional image segmentation methods(binary,watershed,etc.)can only segment single cells or simple overlapping cells,it is difficult to segment highly overlapping cells by traditional image segmentation methods.The segmentation effect is greatly restricted by the quality of original image,which leads to the poor generalization of traditional segmentation methods.Although the cell segmentation generalization ability based on deep learning is strong,it can solve the image segmentation problem of complex background,but the current deep learning method is still difficult to solve the segmentation of highly overlapping cells,and can not achieve the requirement of extracting single cell image from overlapping cells.To solve the above problems,the main works of this thesis are as follows.(1)In order to solve the problem of overlapping cell segmentation,a network model based on optimized Cycle GAN is proposed.The regularization restriction of images is added to the loss function in the model,which fully describes the gap between the generated images and the real images,and accelerates the convergence process.The generator structure in the model is improved by adding skip connection structure,which increases the ability of feature extraction of network.The discriminator uses feature graph instead of scalar value to evaluate the image,so that the calculation of loss function focuses on more details,which ensures the quality of image generation.The experimental results show that the algorithm can effectively generate cell segmentation boundary.(2)In order to ensure the accuracy and rapidity of cell location,Cell?yolo target detection algorithm is proposed in this thesis.Cell?yolo adopts a simplified network structure and improves the maximum pooling operation to retain the information of images to the maximum extent.Aiming at the characteristics of overlapping cells in cervical cell image,a non maximum suppression method of center distance is proposed to prevent the detection frame from being deleted by mistake.At the same time,the loss function is improved and Focal loss is added to alleviate the imbalance between positive and negative samples in the training process.It is verified by experiment,Cell?yolo can ensure the detection accuracy,and the detection rate is 34% higher than YOLOv4.(3)In order to solve the difficulty of boundary attribution of cell images after segmentation and the need of extracting single cell image,this thesis will combine the algorithm of generative adversarial network and target detection to detect and segment,and design a segmentation framework Solo?GAN for extracting individual cell image from cell image,the architecture locks the cell position through target detection algorithm and obtains cell boundary through the generative adversarial networks.The experimental results show that The DC(Dice Coefficient)of Solo?GAN proposed in this thesis is 0.892,which is improved by 23.54% compared with the segmentation algorithm MSSEG based on image processing,and which is improved by 21.52%compared with U-Net segmentation network based on deep learning.The false negative rate of Solo?GAN is 0.163,and the error rate of segmentation is significantly lower than the other two methods.Solo?GAN can segment cell images with high overlap effectively,and has excellent anti-interference ability and meet the actual needs.There are 45 figures,6 tables and 50 references in this thesis.