Research And Application Of Depth Contrast Self-supervised Classification Algorithm For Medical Images

Self-supervised,especially contrastive self-supervised,has attracted wide attention in the medical field because of its characteristics and excellent performance in training unlabeled images.It is difficult to obtain medical image labels,so the combination of self-supervised and medical images has important practical significance.However,The contrastive method directly used for Medical imaging classification may have insufficient model learning ability,which will lead to low accuracy and lack of generalization of the model for downstream datasets..Therefore,the above shortcomings resolved from two aspects: one is multi-directional random erasure data augmentation and contrast reconstruction of self-supervised model;the other is multi-step self-supervised pre-training strategy.The main research contents of this paper are as follows:(1)In order to solve the problem of low accuracy caused by the lack of ability of model learning image features when the contrastive self-supervised method is directly applied to Medical imaging classification,two schemes are proposed in this paper.Firstly,a multi-directional random erasure data augmentation method introduced to optimize the Sim CLR comparison model.Based on the original random erasure method,the area of the erasure region was reduced while the number of the erasure region was increased,so that the erasure area could be spread in all directions,avoiding the obstruction of pathology in medical images due to the large erasure area.Afterwards,this paper proposes the contrast and reconstruction self-supervised model(CRSM)of fusion encoder and decoder.Another self supervised proxy task-image reconstruction introduced in Sim CLR.Based on the encoder of the contrastive model,a corresponding decoder network is constructed to reconstruct the image while performing contrastive learning,thereby improving the classification accuracy of the model.The experimental results show that the model proposed in the paper has good performance and high classification accuracy on the target dataset.(2)In order to further improve the classification accuracy and generalization ability of CRSM on lung disease data sets,this paper proposes a multi-step self-supervised pre-training strategy.Firstly,use the Che Xpert chest disease dataset(including lung diseases)to pre-train CRSM in the first step.Then,initialize CRSM using the model parameters pre-trained in the first step and use it for the second step pre-training of the target dataset(such as Covid-19).Finally,the obtained parameters are migrated to the linear classifier of the downstream task to classify the target dataset.To verify the generalization of the model,in this study,after pre-training the model with Che Xpert in the experiment,its parameters are directly applied to the downstream target dataset classification task,without pre-training the target dataset.Experiments show that the model can be used to classify other lung disease datasets with high accuracy after pre-training of Che Xpert dataset.This strategy can also further improve the classification accuracy of CRSM on target dataset.(3)This paper uses the constructed contrastive and reconstruction selfsupervised model as the backend technical support to design a self-supervised medical image intelligent classification system.Firstly,the user requirements are discussed and analyzed,and the front-end functions including opening image file,image display and displaying result are constructed,as well as the back-end functions of image classification by using the trained model.Through network transmission,the front-end display interface is connected with the back-end model,so that the back-end model classification results can be displayed in the frontend.