Inverse projection group sparse representation for tumor classification: A low rank variation dictionary approach

Abstract

Sparse representation based classification (SRC) achieves good results by addressing recognition problem with sufficient training samples per subject. Tumor classification, however, is a typical small sample problem. In this paper, an inverse projection group sparse representation (IPGSR) model is presented for tumor classification based on constructing a low rank variation dictionary (LRVD), for short, LRVD-IPGSR model. Firstly, an IPGSR model is constructed based on making full use of existing training and test samples, and group sparsity effect of genetic data. Furthermore, from a new viewpoint, a LRVD is constructed for improving the performance of IPGSR-based tumor classification. The LRVD can be independently constructed by detecting and utilizing variations of normals and typical patients, rather than directly using and changed with the genetic data or their corresponding feature data. And the LRVD can be automatic updated and extended to fit the case of new types of diseases. Finally, the LRVD-IPGSR model is fully analyzed from feasibility, stability, optimization and convergence. The performance of the LRVD-IPGSR model-based tumor classification framework is verified on eight microarray gene expression datasets, which contain early diagnosis, tumor type recognition and postoperative metastasis.