Classification Of Medical Data Based On Compressed Sensing

With the promotion of information technology,computer-aided medical diagnosis is widely used.The research shows that computer-aided medical diagnosis technology can reduce the missed diagnosis and improve the diagnostic accuracy.With the improvement of the medical level,DNA microarray technology as an important gene sequencing technology can help us analyze and research disease pathology from the genetic level.It is urgent for us to find a way to dig out effective biological information from massive medical data.At present,the research on the classification of medical data,especially gene expression data,has attracted the attention of many researchers.Some traditional machine learning methods,such as neural network,decision tree and support vector machine,are used to deal with the classification problem of medical data.However,medical data has the characteristics of small samples,high dimensionality and strong noise,traditional methods can not properly deal with the problem.In recent years,compressed sensing as a promising method has been successful in many fields.Aiming at the characteristics that the classification of medical data is difficult,this paper based on the theory of compressed sensing makes an in-depth study on three aspects,which are dictionary learning,sparse representation,multi kernel learning.The research work and main research results are as follows:(1)A classification method based on dictionary learning algorithm(K-SVD)is proposed.In this method,firstly,when the dictionary is updated,we use a group-updated way to optimize the K-SVD algorithm and realize the optimization of the overcomplete dictionary;secondly,in signal reconstruction,we improve the orthogonal matching pursuit algorithm and realize the multiple update at the same time;finally,in the sample classification judgment,the optimized overcomplete dictionary has no longer direct category information,but sparse coefficient matrix obtained by dictionary learning algorithm can be regarded as the sparse representation of training sample set in the optimal dictionary,and the sparse representation of the test sample in the optimal dictionary can be obtained through the reconstruction,so we can use the similarity analysis of sparse representation to classify medical data.This method can optimize the overcomplete dictionary,remove the redundancy of overcomplete dictionary,make the learned dictionary tend to be orthogonal,and solve the small samples problem of medical data effectively.This method not only improves the classification accuracy,but also realizes the feature extraction for second times of medical data after dimensionality reduction.(2)This paper proposes a bidirectional compressive sensing model,then improves the common reconstruction algorithm,and applies it to the solution of the bidirectional compression model and the classification of medical data.The model proposes three common forms:l1 norm,F norm and l2,1 norm,which are analyzed and compared in the reconstruction algorithm and experiment.This model can obtain two-way sparse representation,the sparsity of the row indicates the selection of feature genes,and the sparsity of the column indicates the selection of samples.The solution of this model reduces the redundancy of high dimensional data,removes partial noise and improves the classification accuracy.The obtained row-directional sparse matrix can be used to study the relationship between genes and to provide new knowledge for the study of biological information.(3)This paper proposes a multi kernel learning based compressive sensing model,then improves the common reconstruction algorithm,and applies it to the solution of the multi kernel learning based compressive sensing model and the classification of medical data.This paper introduces kernel method,because the data after kernel operation is related to the number of samples and independent of the dimensionality,kernel method can solve the high dimensionality problem of medical data effectively.The multi kernel learning has better performance,so the idea is introduced.The coarse-grid search method is used to train the training samples to get the optimal kernel combination parameter,and then the test samples are spare represented and classified.The reconstruction method based on multi kernel learning is applied to the classification of medical data,which is helpful to improve the accuracy of classification and represent the similarity between samples.The model can also solve the classification problem of multi-modal data.