Research On Multimodal Assisted Tumor Diagnosis With Machine Learning

In recent years,cancer has brought more and more severe challenges to human health.With the accumulation of medical data and the breakthrough of Artificial Intelligence technology,how to efficiently assist tumor diagnosis has become a challenging problem for Bioinformatics and Computer Scientist.Constructing machine learning models for patients’ different medical data such as clinical,genetic,metabolic and medical images,can help to understand and analysis tumor development from different views,and can efficiently achieve the goal of assisting tumor diagnosis.This paper focuses on studies several key issues in the diagnosis of assisted tumors based on machine learning theory,by analyzing the characteristics of different stages of tumor development and its medical data.The first issue explores how to choose the appropriate Match-Pairs Feature Selection method(MPFS)for screening tumor differentially expressed genes.Only a few genes are differentially expressed during tumorigenesis.Screening out these genes will help to understand the mechanism of tumor development at a deeper level and achieve a more accurate diagnosis of tumors.At present,researchers have obtained a large number of research results in screening differentially expressed genes with feature selection method.However,the MPFS method has not been widely development and research,which considers the case-control pairing characteristics of gene expression data.Therefore,Chapter 3 summarizes MPFS methods in the past ten years,gives its general definition,classifies them into three types,namely Statistical Test type and Conditional Logistic Regression type,and Boosting Strategy class,and finally builds a large number of experiments on the performance and running time of the three types of methods for a comprehensive comparative analysis,for the researchers to choose the appropriate algorithm to provide some reference.The second issue explores how to more accurately screen for tumor differentially expressed genes in matched-pairs data.Tumor tissue contains not only tumor cells,but also other non-tumor cells,and its tumor purity has an important influence on gene differential expression analysis.However,current MPFS methods not consider tumor purity issues in case data when they model the difference in matched-pairs data.Therefore,Chapter 4 proposes a new MPFS method for screening tumor differentially expressed genes,which based on the paired t-test method.The method first estimates tumor purity of case data for each sample,and then estimates the real gene expression of case data,and calculates an optimized paired t-test statistic,and finally screens differentially expressed genes according to a specified threshold value.The experimental results show that the method has high sensitivity and specificity,and the selected genes also have strong biological significance.The third issue explores how to more effectively use medical imaging data to predict gene mutation.After tumor marker genes have been screened,it is important to determine whether or not this gene mutation occurs.Medical imaging is one of the most commonly used diagnostic methods for assisted tumors.It has the advantages of easy access,non-invasiveness,and low cost.Moreover,researchers have found a correlation between image features and gene mutations and began to use medical imaging data to predict whether genes occurred mutation.Current algorithms have the disadvantages of artificial extraction features,two-stage modeling,and the inability to fuse multi-modal medical image data.Therefore,Chapter 5 proposes a Multi-modal 3D Dense Net(M3DDense Net)algorithm for predicting the isocitrate dehydrogenase(IDH)gene mutation in glioma by using medical imaging data.The method uses a 3D convolutional neural network to automatically extract image features,and uses multi-channel technology to fuse multi-modal image information,and achieves the prediction of gene mutations end-to-end.This method has good predictive performance and generalization ability and combines medical imaging and gene data to make the diagnosis of assisted tumor more diversified and reduce the cost of auxiliary diagnosis.The fourth issue explores how to more accurately detect tumor lesions in medical images.The detection of tumor lesions in medical images is the main step in tumor diagnosis,and it is also the basic premise of gene sequencing,gene,and medical image combine analysis,and has important clinical significance.Current algorithms for detecting tumor lesions in mammography is modeled on a single view and does not take into account the fact that tumor lesions are interrelated in both views of the image.Therefore,Chapter 6 proposes a Cross-view Relational Region Convolutional Neural Network(CVR-RCNN)for automatic detection of tumor lesions in mammography.It is the first mammography tumor lesion detection algorithm considering two view imaging information.It uses a two-way object detection architecture to detect lesions in two views simultaneously and proposes a Cross-view Relation Module(CVR)for modeling the relationship between tumor lesions in two view images.This algorithm has higher sensitivity and lower false-positive rate,and can assist imaging doctors to screen tumors,and have strong clinical application value.The main contribution of this paper is conducting related algorithms research from different angles based on the four key issues in assisted tumor diagnosis with machine learning theory: from the perspective of gene,three types MPFS method are compared and analyzed,and proposes a new MPFS considering with tumor purity information;from the perspective of medical imaging,a CVR-RCNN algorithm is proposed to automatically detect tumor lesions in medical images;from the perspective of combining the two,a M3D-Dense Net algorithm is proposed to predict gene mutation by fusing multi-modal medical imaging data.These research work in this paper have a strong frontier,theoretical significance and practical clinical application value,and existing correlation and support between them,which together constitute a preliminary multi-assisted tumor diagnosis system,which for achieving a more accurate multimodal data-assisted tumor diagnosis system in future research work provides a good technical reserve.