| Deep convolution network is widely used in various computer vision tasks because of its excellent ability of learning advanced semantic image features.However,when using convolution network to solve computer vision problems,it usually needs to invest a lot of manually marked image data in the training stage.For the task of medical image segmentation,marking medical image data requires a large number of doctors or professionals to participate,which will lead to the problem of high economic cost.Therefore,in most medical image data sets,there is a problem of insufficient labeled data.Some medical datasets also contain a large number of redundant data without labels.These redundant data can not be directly sent to the network for supervised training.Therefore,it is necessary to use these redundant data reasonably with the help of self supervised learning algorithm.Self-supervised learning is a special unsupervised learning method that data can provide supervision information.It can set up an additional auxiliary task in addition to the conventional supervised learning of neural network.It trains neural network with data without manual annotation in medical image data set,and enables neural network to learn how to extract image visual features more effectively,and solve the problem Better results can be obtained when recording tasks with insufficient samples.This paper takes the chest X-ray image data set disclosed by the third party as the experimental object.Aiming at the defect of the data set with too few segmented and labeled images but a large number of redundant non dimensioned data,combined with the advanced self-monitoring learning algorithm in recent years,the traditional medical image segmentation network U-Net is improved as follows:1)Based on the original U-Net,the paper introduces rotation prediction algorithm to improve the learning ability of the model to the image characteristics,and makes the deep neural network better use the X-ray image data without segmentation label in the data set to understand the position and posture of the instance object in the image.2)In order to improve the image feature extraction ability of the first half of the U-Net network,the paper chooses to apply the maximum mutual information algorithm to the encoder.Using the chest X-ray image data without label,the encoder is trained separately,so that the encoder can get better image feature extraction ability.3)The segmentation performance of convolution network with self supervised learning is very dependent on training strategy.Therefore,while introducing the algorithm,this paper also designs the corresponding algorithm flow framework and training strategy.This paper verifies the proposed algorithm model by many experiments,and evaluates the output results of the algorithm by using the four indexes,i.e.Miu value,dice similarity coefficient,sensitivity and accuracy.Firstly,the algorithm of this paper is used to perform lung tissue segmentation experiment on X-ray images,and compare the results with the U-Net,FCN,Seg Net model and XNet in the same dataset under the same experimental environment,which proves that this method is more suitable for Xray image segmentation.Secondly,experiments are set to verify the value of the weight parameter in the loss function.Thirdly,in order to verify the generality of the algorithm in this paper,a number of comparative experiments are carried out using liver CT dataset.Fourthly,we evaluate and compare the time-consuming of our algorithm and other segmentation algorithms in training and testing,and put forward the next work suggestions.The experiment shows that the use of rotation prediction algorithm and maximum mutual information algorithm can improve the accuracy of u-net model in medical image segmentation. |
