Research On The Semi-supervised Medical Image Classification Based On Pseudo-labeling And Consistency Regularization Methods

Deep learning has achieved remarkable results in medical image assisted diagnosis,especially in the field of medical image classification.However,it is expensive to build large-scale labeled medical image datasets,and semi-supervised deep learning models using a small amount of labeled data and a large amount of unlabeled data for network training have become a research hotspot in the field of medical image classification under the current conditions of difficult access to labeled medical data.However,the mainstream methods in this field still have the following problems that need to be solved: first,the pseudo-labeling method filters out a large amount of data below a set threshold,making the utilization of unlabeled data low;the consistent regularization method relies heavily on the perturbation function and has the defect of insufficient generalization ability.Second,the medical image dataset has a serious class imbalance problem,which can lead to model prediction biased toward the head class;the consistency regularization method usually ignores the problem that the reliability of unlabeled medical images acquired in clinical practice varies due to different environmental factors.Aiming at problem one,this paper proposes a semi-supervised medical image classification algorithm based on class prototype matching for soft pseudo-labeling with consistency regularization is proposed.First,a class prototype matching module is used to predict soft pseudo labels of unlabeled data,and the quality of the predicted pseudo labels is improved by an elaborate dynamic and unbiased update cache queue.Then,more complex data inputs are provided to the model by linearly mixing labeled and unlabeled data and their corresponding labels to improve the model’s ability to learn intra-class and inter-class features.The prediction accuracy of the model is improved by constraining the model to always maintain consistent predictions for unlabeled samples through the consistency regularization term of unlabeled data under different augmentation methods.A series of comparative and ablation experiments on the ISIC2018 skin lesion dataset and the CheXpert chest disease dataset showed that this algorithm effectively incorporates the advantages of pseudolabeling and consistency regularization methods,and that its classification performance and generalization ability outperform other state-of-the-art semisupervised classification methods.Aiming at problem two,this paper proposes a semi-supervised medical image classification algorithm based on adaptive confidence pseudo labeling with weighted consistency regularization.The method generates pseudo-labels by adaptively adjusting the thresholds for each class,and reduces the thresholds for classes with poor learning status to encourage the model to learn more unlabeled data belonging to that class,thus reducing the prediction bias caused by the imbalance problem of medical image data.In addition,a weight evaluation model is introduced to map the prediction distribution of unlabeled data to weights reflecting their corresponding reliability in order to improve the model’s ability to predict consistently for unlabeled data containing different reliability.Experimental results show that this method can effectively reduce the prediction bias of the model in the case of class imbalance,and the classification performance on the ISIC2018 skin lesion dataset and the BCCD blood cell dataset outperforms other state-of-the-art semi-supervised methods.