Discussion On Symmetry Network In Biomedical Image Segmentation

Biomedical image segmentation has always been a challenging task.In recent years,how to solve related problems in the field of biomedicine through deep learning technology has become a research topic.A large number of studies have proved that the current convolutional neural networks(CNN)can solve many segmentation tasks,but due to the sensitivity of biomedical images,it is difficult to obtain a large number of labeled datasets,which increases the difficulty of model training.In addition,in order to obtain a better segmentation accuracy,a complex neural network model is chosen,but this would reduce the training speed and take up much memory space.Therefore,building a network that is efficient,high-precision,and applicable to some small-scale biomedical image datasets is the key to overcome this problem.This paper discusses and designs a symmetrical deep learning network model to solve the problem of biomedical image segmentation.The main work and results are as follows:(1)In view of the small scale of biomedical image datasets,which leads to insufficient feature information,this paper adopts efficient and simple data enhancement methods to expand the image datasets through rotation transformation,and increases the size of the datasets.(2)Aiming at the butterfly effect of the Dice loss function in the model training process,which leads to the instability of the model training,it is difficult to find the optimal solution.This article starts from the partial derivative of the Dice loss function in the backpropagation and analyzes its gradient from a numerical perspective to find the cause of the butterfly effect.And the log function is used to optimize the Dice loss to alleviate the butterfly effect generated during the training process,making the model training more stable and easier to find the optimal solution.(3)To discuss the classic deep learning model U-Net in the field of biomedical imaging,in order to reduce the number of parameters,speed up training,and increase the accuracy of segmentation,the U-Net symmetry structure is used as the basic skeleton of the network model designed in this paper.On this basis,the concepts of Dense-Net,Res-Net and Inception-Net are introduced,and they are assembled into a symmetric network model in the form of functional modules,which are optimized from two aspects of increasing the receptive field and reusing feature information.In addition,the number of model layers is explored to find a shallow network backbone.It aims to construct a rapid and symmetric network model suitable for biomedical images.This paper compares the training results between the classic symmetric network U-Net equipped with Dice loss and the optimized Dice loss on the public biomedical image datasets.According to the obtained training curve,the optimized Dice loss has a smoother curve.Experiments prove that the optimized Dice loss is effective for mitigating the butterfly effect caused by Dice loss.On this basis,the optimized Dice is used as the loss function.Compared to U-Net,the performance of the MIRD-Net model and the U-Net-12 + IT-Block(3,12)model are both improved.Moreover,the 3D-MIRD-NET,which is extended from MIRD-Net,outperforms the 3D-UNet in segmentation accuracy under the condition of lower parameter and faster training speed.The experimental results show that in the task of biomedical image segmentation,designing a stable loss function,building a specific network skeleton,and embedding other classic networks as modularization are of positive significance for optimizing deep neural network models.