Feature Representation And Inverse Projection Sparse Representation Classification For Clinical Medical Data Analysis

Medical clinical diagnosis is closely related to human health.With the rapid development of science and technology,the wide application of advanced medical equipment,and the effective storage of clinical medical data of a large number of patients,the demand for discovering effective information contained in medical data is increasingly urgent.However,clinical data often manifests as diversity,high dimensionality,heterogeneity and redundancy,which pose a huge challenge to traditional clinical diagnostic methods.Exploring reasonable data analysis method has great significance for effective processing and analyzing of medical data,assisting doctors to diagnosis,guiding clinical treatments and preventing diseases.During the diagnosis and prognostic analysis of disease,there are many clinical influencing factors.Doctors pay attention to the screening of important influencing factors.For the medical image data,some features are invisible or difficult to find.In addition,there is a large amount of unlabeled data in medical databases,and then it is important to fully mining the data information.Based on the characteristics of clinical data,the efficient data analysis needs to solve two main problems: one is the feature representation,which is critical to analysis and process medical data.The other is effective classification,where the good classification algorithms are designed to achieve the purposes of accurate identification and prediction.The main work of this thesis is summarized as follows:(1)A clinically important factor analysis method is proposed based on decision curve analysis and inverse projection sparse representation classification.Firstly,for the analysis of important influencing factors of clinical prognosis,on the feature representation,an automatic screening method of important influencing factors is given based on decision curve analysis,and the important influencing factors are sorted.This method is not only suitable for continuous data,but also effective for discrete data.By independent analysis of every factor,it not only effectively avoids the problem of dumb variables and multi-factor multi-collinearity in traditional medical analysis methods,but also effectively completes screening of important influencing factors(i.e.feature selection).Second,based on this feature selection,the prognostic analysis is completed by inverse projection sparse representation classification,which is an improvement of sparse representation based classification method.Inverse projection sparse representationclassification not only effectively overcomes the excessive dependence on sufficient training samples,but also fully exploits the information contained in the test samples.And it also reduces the effect of the number of training samples on the prediction results.The clinical data provided by Chongqing Medical University has been tested to verify that the proposed method is very consistent with the clinician's diagnosis.(2)A tumor recognition method is proposed based on deep random forest and inverse projection sparse representation classification.For the medical image data,the effective training samples with annotations are few and difficult to obtain.In the feature representation,this thesis introduces the deep forest method to fulfill deep feature learning.Then,the inverse projection sparse representation classification method is combined with deep forest-based feature learning,and the classification is completed.On the CT dataset of clinical esophageal cancer provided by Henan Cancer Hospital,the characteristics of the esophageal tumor CT data of single patient and multiple patients have been studied and classified.Experimental results verify the feasibility and effectiveness of the proposed method.