Research On Medical Image Segmentation Algorithm Based On Convolutional Neural Network

Medical image segmentation is a key step in the field of medical image processing and analysis.It is considered as basis of the detection and recognition of lesions,and provides a theoretical basis for the diagnosis and analysis of diseases and subsequent treatment.In recent years,with the rapid development of deep learning,medical image processing based on convolutional neural network has become a research hotspot,which is the most widely accepted method at present.However,local operation of convolutional neural network limits the development of medical image segmentation tasks,so more efficient segmentation algorithms need to be explored.In this paper,due to the small number of medical images in the data set,low contrast,small targets to be segmented,irregular shape,and the blurred boundaries,the segmentation results are not ideal.The following research work is carried out:(1)In order to solve the problems that traditional deep learning network model has too many parameters and is difficult to be applied in actual clinical practice,this paper proposes a new coronavirus lesion segmentation model based on lightweight context information fusion network.For segmentation model,double encoders and decoders are used,and global pyramid guidance module and scale-aware pyramid fusion module are introduced to obtain better fuse shallow and deep semantic information,which captures the feature information more precise and improves the segmentation performance of the network model.The segmentation model is validated on two COVID-19 datasets,and the experimental results show that the number of model parameters is about1.68 M,which is greatly reduced compared with the traditional U-Net model and the segmentation performance is better than other segmentation methods.(2)In order to solve the problem of a few labels in medical images,this paper proposes a medical image segmentation model based on regularization-driven semi-supervised learning.Firstly,the model includes student model and teacher model,which make full use of labeled and unlabeled images to make them learn and promote each other in the training process.Secondly,the virtual adversarial training strategy is introduced to make the unlabeled images more sensitive to small disturbances and improve the robustness of the model.Then,the entropy minimization principle is used to balance the decision boundary effectively by adding entropy to the unlabeled images.Finally,the bayesian optimization method is applied to automatically search the optimal combination of different hyper-parameters in virtual adversarial training,which improves the segmentation performance of the model.The segmentation model is validated on skin cancer ISIC2017 and COVID-19 CT datasets.Experimental results show that the proposed method improves the segmentation performance and is superior to other semi-supervised segmentation methods.