Research On Key Technologies Of Medical Image Classification Based On Unsupervised And Semi-supervised Framework

In recent years,computer-aided diagnosis technology shows a great vitality in clinical work,accelerated by the rapid development of deep learning.As one of the most widely used means in computer-aided diagnosis,medical image classification can realize the screening of diseases and the classification of lesions,which is of great significance for human health and well-being.However,most exiting medical image classification methods rely on the supervised framework with completely labeled data,which requires large amount of manual labor on the medical image annotation.In addition,it is hard to avoid the noisy annotations even though an experienced expertize.This undoubtedly brings plenty of limitations to the research work.The medical images classification methods based on unsupervised and semisupervised framework can break through the research dilemma caused by the scarcity of labeled medical image data,while these research works are still confronted with several weaknesses and bottlenecks.This thesis carries out the key technology research of unsupervised and semi-supervised medical image classification to solve the problem of insufficient labeled medical image data.Aiming at several specific medical image classification tasks with randomness and specificity,this thesis explores effective models to improve the diagnostic performance of medical image classification.The main research contents and innovative achievements of this thesis are summarized below:1.This thesis proposes a Self-supervised Topology Clustering Network(STCN)for unsupervised medical image classification,which builds a self-supervised transformation-invariant network to settle the uncertainty and complexity of medical images caused by the variances of illumination,contrast,view point,and background.This network can normalize the images under changing clinical circumstances to enhance the robustness of the learned feature representations.To exploit the topology relationship in unlabeled medical images,STCN introduces a deep topology clustering module to automatically partition the cluster groups without a prior-acknowledged category number,inspired by the modularity maximum algorithm.Finally,STCN is optimized through a self-supervised training process to improve the classification performance.This method can effectively settle the unlabeled medical image classification problem without any prior information(such as category number),and it obtains a satisfied performance on unsupervised skin cancer images classification task.2.This thesis proposes a Prototype Transfer Generative Adversarial Network(PTGAN)to solve the unsupervised breast tumor images classification task.According to the difference in style,distribution of breast tumor images which are always captured by different devices and parameters,the PTGAN designs a target-biased generative adversarial network to alleviate the pixel-level style gap between different datasets by the adversarial learning between domain discriminator and generator.To leverage the lacking of labeled data,PTGAN builds the prototype transfer module which can further transfer the knowledge in a completely labeled source domain into the unlabeled target domain by the cross domain loss.Then,PTGAN calculates the pseudo labels for the unlabeled target data by the prototype classifier,and bridges the class-level gap between source and target domains.This method obtains an average classification accuracy of 87.6% on unlabeled breast tumor images classification with different magnification scales,which demonstrates its commendable scalability.3.This thesis proposes a Multi-source Fuzzy Attention Network(MFAN)to solve the semi-supervised breast tumor image classification problem.MFAN designs a domain-attention feature extractor and domain-invariant generator aiming at the scarcity of labeled data and the poor performance of unsupervised framework.It utilizes a handful of labeled data and plenty of unlabeled images to learn a breast tumor image self-reconstruction model,with the help of multiple labeled source datasets.Considering the different contributions of different domains,MFAN introduces fuzzy clustering algorithm to exploit the weights of each source and transforms images into a target-biased feature space,which improves the robustness by the adversarial learning between image decoder and realistic discriminator.Finally,MFAN employs K-means to reason the pseudo labels for target data,which are utilized to train a classification model on the learned feature space.From extensive experiments,MFAN method can promote the classification performance on semi-supervised medical images classification task,which benefits the clinical screening of breast tumor.4.This thesis proposes a Deep Multi-feature Reinforcement Network(DMRN)for semi-supervised brain image classification problem.In order to solve the problem that there are few labeled data and poor interpretability of deep learning,DMRN designs the feature extractor module to learn deep representations,rotation-invariant and grayinvariant texture feature,and generalized linear morphology feature.DMRN fuses the deep representation and those two hand-craft features in the multi-feature reinforcement module,which provides the synthetic feature to the deconvolutional reconstruction network to improve the representative ability for images.Finally,DMRN trains a classification model for the fused feature representations and obtains performable results,compared to existing semi-supervised brain image classification methods,where remains a limited distance between supervised methods.