HEp-2 Cells Staining Pattern Classification With Local Texture Features Description

Analysis of antinuclear antibodies(ANA) by indirect immunefluorescence(IIF) imaging is an effective means of modern medical diagnosis of autoimmune diseases, since the large-scale detection and analysis is a very tedious and time-consuming work, and in order to improve the detection of efficiency and reliability, automatic ANA analysis system based on computer vision is badly needed to aid specialist for diagnosis.At present, IIF fluorescence image of analysis focuses on the staining pattern classification, different staining patterns are closely related to various autoimmune diseases, and therefore the staining pattern classification accuracy is important for the final diagnosis. Experts classified the current pattern of staining mainly six categories, i.e, Centromere, Homogeneous, Nucleolar, Coarse Speckled, Fine Speckled and Cytoplasmic. The accurate classification of staining pattern is still a very difficult problem, the reasons are that the environment of fluorescence imaging has relatively large differences result in the great appearance variation of different cell staining pattern images, and some of them also have a very similar visual characteristics, therefor, how to obtain higher robust feature to describe the staining pattern characteristic is still a very difficult problem.Fluorescent staining pattern classification primarily includes image pre-processing, image feature extraction, classifier design, among them, compared to the pre-processing and classifier, feature extraction algorithm is the key technology for success, The focus of this paper is to extract information adaptable to describe the staining pattern characteristics, according to the literature and the study of staining patterns, a robust texture feature is more effectively to describe the staining pattern characteristics, Therefore, this work consider local texture, i.e., LBP and shape index feature for HEp-2 staining pattern classification. And they are used to the ICPR 2012 contest data sets and SNP HEp-2 datasets experiments. The main works of the paper are as follows:1) The LBP and improved CLBP features are used in HEP-2 staining pattern classification, moreover a novel neural network ensemble classifier and the classic LIBSVM classifier are adopted to verify the effectiveness of LBP features.2) Based on the description ability for second-order image structure with shape index, a texture descriptor adapt to HEp-2 cell image is introduced, moreover, combining scale space theory and intensity space decomposition strategy, a novel feature that is suitable for HEp-2 staining pattern classification is proposed. This feature has the ability to describe second-order image structure and spatial information. The method achieved very good classification results in the ICPR 2012 contest data sets and SNP HEp-2 data set. Experimental results show that the proposed method is superior to other popular texture descriptors, such as: LBP,CLBP. and approximate to the performance of CoALBP feature, which is the first winner of the 2012 ICPR contest.