Pathological Image Diagnosis Of Children’s Tumors Based On Deep Learning

In recent years,a lot of work has proved the superiority of artificial intelligence in many fields,such as image recognition,natural language processing,speech recognition,etc.The technology can greatly reduce repetitive labor costs.Thanks to the development of medical image digitization technology,artificial intelligence-assisted medical diagnosis has become a possibility.The current attempts of deep learning technology in adult pathology tasks are common,but the application in pediatric tumor pathology tasks is relatively rare.Childhood cancer is a leading cause of children death.The world is currently facing a serious shortage of experienced pediatric pathologists.Therefore,it is urgent to develop artificial intelligence algorithms that can assist pediatric pathologists in their daily diagnosis.When using deep learning to assist the pathological diagnosis of children’s tumors,three crucial problems need to be solved,namely: the problem of key feature extraction? the problem of small samples? the problem of fusion of diverse information sources.Around the three parallel problems,we propose targeted solutions separately while focusing on the improvement of model interpretability.Compared with other types of images,pathological images have many unique features,so efficient directional extraction of pathological image features is an important issue.Pathological images do not have a clear distinction between image objects and backgrounds like natural images,and the purpose of pathological image analysis is often not to extract objects or analyze higher-level semantic features,but to analyze cell morphology,and arrangement of tissues or cells,etc.,and then draw the conclusion.Therefore,in the pathological image analysis task,some special designs for models should be applied according to the characteristics of the pathological image.For the task of classification with seven categories（six sub-categories in neuroblastoma and one normal cell category）,we propose to combine neural network with TEM（texture energy measure）,and propose a novel network architecture named Detex Net（deep texture network）,which is the first contribution.By introducing expert knowledge as a priori at the bottom layer of the network,this method makes the meaning of the underlying representation pattern clearer,so that the network can capture the key information of pathological images more smoothly.The model achieves performance improvement by directional capture of key information,and provides ideas worthy of reference for research tasks with obvious field characteristics.Due to the low incidence of childhood cancer,the small population base of children,and the difficulty in obtaining relevant slice data,the small sample problem is more prominent in the field of pediatric tumor pathology research.High noise and high diversity are the main reasons for poor model performance in small sample scenarios.We propose an innovative model Saga Net（small sample gated network）to solve the binary diagnosis task of SRBCTs（small round blue cell tumors）,which is our second contribution.We actively mask noise sources and make the model focus on valid cell regions in pathological images by designing a mask filtering mechanism,and propose a length-aware hinge loss to improve the model’s tolerance to feature diversity.Experiments demonstrate that our method achieves significant gains and outperforms current mainstream models.This study shows that it is feasible and necessary to actively block useless information when the amount of data is obviously insufficient.Moreover,it is an effective approach to enhance the model’s ability to distinguish features by adding explicit mapping relationships to the loss function.The role of diverse information sources in medical data for specific tasks is often unclear.Therefore,how to comprehensively utilize this information and achieve complementarity between different information sources is an important issue worthy of attention.For a four-class classification task with three subcategories and one normal cell category in neuroblastoma,we propose to use a multi-view maximum entropy discriminant（MVMED）model to solve this problem,the third contribution.We utilize interval consistency theory and use variational optimization under a Bayesian learning framework to obtain the posterior distribution of the classifier.In the experiments,the multi-view model outperforms single-view and other mainstream models,and it is proved that the automatic feature selection by the multi-view learning method with shared classification intervals is suitable for the task scenarios of different information sources.The fourth contribution is to supplement the target task with information by introducing external information sources in different forms of data.In the automatic classification task of protein detection sequences for SRBCTs,based on the fact that pathologists diagnose sequences according to protein functions,we propose to construct a knowledge graph of commonly used pediatric proteins in the upstream task and map them into embeddings.The obtained protein function embeddings are used for downstream protein detection sequence classification tasks.The experimental results demonstrate that the information of the upstream task is effectively fused into the information of the downstream task and the proposed model achieves the best classification performance.This strategy of selecting the category of information captured by the upstream task through the discriminant logic of the downstream task is also applicable to other researches with special data forms and insufficient information.High interpretability is a necessary condition for the application of deep learning in the medical field,so we focus on the improvement of interpretability while solving the corresponding problems in each work.In the first work,we extract texture features by specifying the network to make the feature meaning more explicit? in the second work,we filter out image noise in an explicit way? in the third work,we extract different types of traditional features for multi-view learning,so that image features are rich in information without losing clear physical meaning? in the fourth work,we explicitly construct a protein knowledge graph and transparentize the decision-making process in downstream tasks by visualizing the self-attention layer and proposing two functional measures of protein importance and coupling.