Low Rank Variation Dictionary And Inverse-projection Group Sparse Representation Model For Tumor Classification

Malignant tumor is a major disease that seriously harms human life and health.With the rapid development of bioinformatics,the use of microarray gene expression data to study the tumorigenesis and development mechanism at the gene level is helpful for tumor diagnosis and personalized treatment.Sparse representation(SR)and low rank decomposition are the key methods for tumor classification based on microarray gene expression data.Sparse representation classification achieves good results by addressing recognition problem with sufficient training samples per subject.However,SR classification performs not very well for small sample data.Inverse projection-based pseudo-full-space representation classification(PFSRC)is an improved improved SR classification method.PFSRC focused on exploiting complementary information between training samples and test samples by utilizing existing available samples.However,the microarray gene expression data does not have such complementarity.The low-rank decomposition decomposes the data into low-rank part and sparse part.And using the low-rank part of the information that reflects the common information between samples restore the original sample information,however,the effect of sparse part reflecting the lesion information between samples on tumor classification is neglected.Based on this,this article mainly did the following research work:(1)Two kinds of low-rank variation dictionary are constructed.From a new viewpoint,tumor recognition problem is tackled from detecting and utilizing variations in microarray gene expression data of normals and patients,rather than directly using these gene data.Based on this,two LR variation dictionaries(fixed and changed elements)are constructed according to actual clinical needs.(2)The inverse projection group sparsity representation(IPGSR)model is constructed based on making full use of existing samples and group effect of microarray gene data.IPGSR stability is analyzed.And then,a classification criterion named category contribution rate is constructed to match inverse projection group sparsity representation and complete the IPGSR classification.Furthermore,the optimization and convergence of IPGSR model is analyzed and proved.(3)Based on the theory of low rank variation dictionary and inverse-projection group sparse representation,a low rank variation dictionary and inverse-projection group sparse representation model for tumor classification is proposed.Extensive experiments on the seven public tumor microarray gene expression datasets demonstrate the proposed technique is competitive with state-of-the-art methods.Finally,biology analysis of candidate pathogenic genes is given.