Research On Medical Image Classification Algorithm Based On Interpretable Deep Learning And Self-supervised Learning

In recent years,the rapid development of deep learning has greatly promoted the scientific research process of researchers in natural language processing,computer vision,and other fields.The deep learning technologies have rapidly been applied to medical care,transportation,finance,and other fields.With the research of the interpretable models and self-supervised learning,the application of deep learning technology in the medical field has obtained significant development and achieved good results.Image classification in the medical field is different from other fields.On the one hand,due to the particularity of the medical industry,the relatively small number of datasets and limited data annotations limit the development of medical image classification algorithms.On the other hand,in the medical image classification tasks,the algorithm must provide interpretability,and the medical diagnosis process must be interpretable and transparent.Therefore,the research of medical image classification algorithms needs to pay attention to the characteristics of "small data,large task" and the importance of interpretability.Based on this,this thesis proposes a medical image classification algorithm based on interpretable deep learning and self-supervised learning to solve the above problems.For the requirement of interpretability of medical image classification,this thesis proposes a fine-grained image classification interpretable model MBC＿Proto Tree based on an improved neural prototype tree.The model can provide global interpretability and local interpretability for the classification decision process.A new multi-granularity feature extraction network is designed,and three new backbone networks are used for feature extraction,which improves the feature extraction performance of fine-grained and multi-granularity images.A background prototype removal mechanism is designed in the soft neural binary decision tree layer,which ensures the correct update of the prototype and realizes the optimization of the prototype path decision.A new loss function with both a leaf node loss function and a fully connected layer loss function is designed to improve the generalization ability of the model.We conducted comparative experiments on the fine-grained datasets CUB-200-2011 and FGVC-Aircraft.Finally,a comparative experiment was performed specifically on Chest X-Ray medical image dataset.The experiment results show that MBC＿Proto Tree algorithm not only improves the accuracy of image classification,but also provides better interpretation.For the problem of "small data,big task" caused by the high cost of data annotation by experts in the field of medical images,this thesis proposes a self-supervised comparative learning medical image classification algorithm HPSCM based on hierarchical prototype tree.The model only needs a small number of image labels to finetune,and can obtain a model with performance comparable to supervised learning,and provide an interpretable basis for the classification decision process.The model first designs a three-encoder for feature extraction,so as to ensure that the encoder saves more information in the learned representation.Afterwards,a dual mode of instance contrast learning and prototype contrast learning using rows and columns of feature matrices is designed to solve the influence of image noise information on contrast learning,more correct semantic information can be used in contrast learning.Furthermore,we designed a prototype comparison selection module based on hierarchical prototype tree,introduced neural prototype tree on the basis of prototype clustering,optimized the process of prototype training,ensured the model to make better prototype decision,and improved the interpretability of prototype decision.We pre-trained and performed linear classification evaluation,KNN classification evaluation,semi-supervised learning evaluation on Image Net dataset,then we performed transfer learning experiments on Places205,VOC07.Finally,we performed medical image classification experiment under different downstream label fractions on Chest14,Che Xpert.Our model achieve preferable results on these datasets.In this thesis,aiming at the problems of interpretable deep learning model and corresponding self-supervised learning model in medical image classification algorithm,the algorithm innovation is carried out respectively,and the corresponding algorithm model is proposed.Experiments show that the above two algorithms achieve advanced performance in interpreting fine-grained medical image classification task and selfsupervised contrast learning medical image classification tasks,respectively.