HEp-2 Cell Classification Using Multi-scale Texture Information and Multiple Kernel Learning

Indirect immunofluorescence imaging is employed as a standard method to detect antinuclear antibodies in HEp-2 cells which is important for diagnosing autoimmune diseases and other important pathological conditions involving the immune system. HEp-2 cells are generally categorised into six groups: homogeneous, fine speckled, coarse speckled, nucleolar, cytoplasmic, and centromere cells, which give indications on different autoimmune diseases. Typically, this categorisation is performed manually by an expert and is consequently a time consuming and subjective task. In this paper, we present a method for automatically classifiying HEp-2 cells using multi-scale texture descriptors and multiple kernel learning based classification. We extract local binary pattern (LBP) texture features and summarise these in form of multi-dimensional LBP (MD-LBP) histograms to maintain the relationships between the scales. We then employ a multiple kernel based approach using different support vector machines with polynomial kernels for classification. We evaluate our algorithm on the ICPR 2012 HEp-2 contest benchmark dataset, and demonstrate it to outperform all but one of the algorithms that were entered in the competition.