Fisher Vector of Micro-Texton for HEp-2 Staining Pattern Classification

This study addresses the classification problem of HEp-2 cell using indirect immunofluorescent (IIF) image analysis, which can indicate the presence of autoimmune diseases by searching for antibodies in the patient serum. Generally, IIF analysis remains a subjective method, which depends too heavily on the experience and expertise of the physician. Recently, some studies show that it is possible to identify the cell patterns using IIF by image analysis and machine learning techniques. However, it still has large gap between automatic recognition and the physical experts’ decision. This paper explores the discriminative feature extraction of HEp-2 cell images in IIF, and then identifies the patterns of HEp-2 cell using machine learning techniques. Motivated by the research progress on computer vision that small local pixel pattern distributions can be highly discriminative, the proposed strategy employs a parametric probability process to model the local image patches (Textons: micro structure in the cell image), and extract the higher-order statistics (also called Fisher-Vector) to the model parameters for the image description. The proposed strategy can adaptively characterize the micro-Texton space of HEp-2 cell images as the generative probability model, and learn the parameters for better fitting the training space, which would lead to more discriminant representation for the cell image. The simple linear support vector machine is combined for cell pattern identification due to its low computational cost especially for large-scale dataset. Experiments on the released HEp-2 cell dataset of ICIP2013 competition validate that the proposed strategy can achieve much better performance than the popular used local binary pattern (LBP) image descriptor, and the achieved recognition error rate is even greatly below the observed intra-laboratory variability.