Benchmarking HEp-2 specimen cells classification using linear discriminant analysis on higher order spectra features of cell shape

Abstract

Human Epithelial Type-2 (HEp-2) specimen patterns in microscopic images have been widely used for diagnosing autoimmune diseases. Manual cell classification is inefficient, time consuming and subject to human error, therefore, researchers have attempted to automate the procedure of HEp-2 cell classification. Challenges include large data volumes, variation in cell densities, shape, rotation, scaling and shifting, and overlapped cells. A computer-aided diagnosis (CAD) system, based on image-based classification, can help in terms of effort, time, and reliability of classification HEp-2 specimens. HEp-2 cells are classified into seven classes: speckled, homogeneous, centromere, nuclear membrane, nucleolar, golgi, and mitotic spindle. In this work, we focus on the task of HEp-2 specimen cell classification in terms of feature extraction representations, features selection and classifier design. We propose a method of linear discriminant analysis relying on bispectral invariant features from a segmented HEp-2 specimen cell shape that allows us to generalize features to being variation in shape, rotation, scaling and shifting. HEp-2 specimen cell shape segmentation is determined using level set based forces in normal direction and geometric active contours. Both linear and kernel support vector machines as well as fuzzy C mean are employed for fast classification of specimens and achieved Mean Class Rates (MCAs) of 91.25%, 86.02% and 89.14%, respectively, over the ICPR2014 and ICPR2016 contests HEp-2 cell task-2 datasets. The results validate the competitive performance of the proposed system compared with other approaches reported on both contests.