Research On Prediction Methods Of Digital Pathological Images Based On Capsule Neural Network

With the increase in morbidity and mortality of cancer,accurate diagnosis of cancer patients is particularly important.Because the pathological images contain a lot of clinical information closely related to patients,which is significant for the diagnosis of cancer.Therefore,accurate prediction of pathological images can assist doctors to make accurate clinical diagnosis and predict the development of the disease,which provides great help in making appropriate treatment protocols and improving the quality of patients’treatment.In recent years,with the development of automated tissue sectioning and whole slide scanning technology,a large number of cancer digital pathological images have been accumulated.Meanwhile,thanks to the rapid development of artificial intelligence technology,the use of artificial intelligence technology can greatly improve the prediction accuracy of cancer digital pathology images.Therefore,it is urgent to propose new prediction methods based on artificial intelligence technology for accurate digital pathological image prediction,which can assist doctors in diagnosis and treatment.At present,most of the artificial intelligence methods for pathological image prediction are based on convolutional neural networks.So,the spatial and hierarchical relationships between objects cannot be considered,which has great limitations.In order to solve the above problems,this paper uses capsule network,an emerging artificial intelligence method,to propose new prediction methods for two important tasks of digital pathological image prediction.The main contributions of this paper are as follows:(1)A novel classification method named CapCell is proposed in this study.CapCell is based on capsule network and is aimed at predicting the type of cells according to cell pathological images.Meanwhile,by combining two different loss functions,a novel loss function named cell loss is designed for CapCell,which can help accurately classify lymphocytes,normal epithelial cells and malignant epithelial cells of breast cancer.Then,a series of evaluation metrics such as accuracy are used to evaluate the prediction performance of CapCell in independent test set.Compared with other classification methods,CapCell shows its excellent prediction performance.Finally,by analyzing the weight of cell loss,the effectiveness of CapCell is also verified(2)In order to predict the survival time of patients by using whole slide pathological images,a new variant of capsule network named CapSurv is proposed in this study.By integrating cox loss,margin loss and reconstruction loss,we design a new loss function called survival loss specially targeting to conduct survival analysis for CapSurv.Besides,to optimize the training of CapSurv,we collect discriminative patches closely relevant to survival time by extracting features of patches utilizing VGKG16 network and clustering.Then,by using multiple performance metrics,we provide a comprehensive assessment of the prediction performance of CapSurv in two different cancer datasets.Meanwhile,the comparison with other survival models shows the superior perform ance of CapSurv.Finally,the pathological patch screening method and survival loss are analyzed and evaluated respectively.The results verify the effectiveness of CapSurv in survival prediction with whole slide pathological images.