| In the diagnosis and treatment of COVID-19,professional physicians judge the patient’s condition by observing the condition of the patient’s lung lesions,which is similar to the medical image segmentation task that label is called segmentation map,and the purpose is to extract areas of interest in the target image,such as lesions,tissues organs,etc.,due to a lot of CT images and the large workload of professional physicians,it is of great significance to use computer-aided physicians for diagnosis and treatment.Medical images often lack standard data.Semi-supervised learning can reduce reliance on standard data.Some semi-supervised learning algorithms obtain high-quality pseudolabels by consistency regularization,but don’t consider the uncertainty of estimating pseudo-labels;co-training is a semi-supervised learning method,and the problem of standard co-training is that if the quality of each label is poor,then the training performance of the base network will decrease.To sum up,we propose an approach from co-training:(1)Analyze the deficiencies of standard co-training methods,and a novel semi-supervised coronavirus pneumonia lesion segmentation approach on lung CT images based on uncertainty estimation is proposed.(2)In our approach,it is proposed to alleviate the problem of low quality of pseudo-labels in standard co-training by ensemble learning,meanwhile,a way to dynamically calculate the weighting coefficient according to the training loss between the basic networks is proposed.(3)Generally,it is necessary to define a threshold to convert the probability distribution of the network output into pseudo-labels.In the case of automatic threshold to generate pseudo-labels,we use JS distance to measure the uncertainty of pseudo-labels and improves the accuracy of pseudo-labels supervised loss.(4)Experiments on public and private datasets,compared with existing semi-supervised methods,have improved in both similarity-based and confusion-matrix-based metrics. |
PDF Full Text Request