In this chapter, we have presented a promising framework, AdaCoDT, for automatic staining pattern classification of HEp-2 cells to support the diagnosis of specific autoimmune diseases. It is verified that LBP related features can be successfully applied for classifying the staining pattern of HEp-2 cells based on the recent researches on the staining patterns analysis and classification. However, the LBP quantizes the local structures into only two quantized levels by thresholding the neighboring pixel intensity with that of the center one, thus some important information is lost. To reserve more discriminative information, we propose a CoDT feature which directly adpots the differential vectors of micro-texton and its neighborhoods. It further captures the spatial information between neighboring micro-textons, thus it can provide strong discriminative and descriptive capability.
In addition, the BoW framework, which is one of the most famous and efficient approaches for image categorization, has been applied in classification of HEp-2 staining patterns and has shown impressive performance. Unfortunately, the BoW approach suffers from its own problems which are the inevitable information loss in feature quantization process and the high computational cost to built BoW representation. To handle these problems, we make use of the FK principal for staining pattern classification. It characterizes an image by a gradient vector derived from a generative process of the training data, thus the resulting representation is adapted to the staining pattern classification task. Meanwhile, with the same size of vocabulary, the FK based method can extract much larger image representation than the BoW representation. Hence, it can provide excellent performance with simple linear classifier.
Within the proposed AdaCoDT framework, we first extract the CoDT feature from each HEp-2 cell image. Then, we approximate the distribution of CoDT feature as a GMM which can adaptively partition the CoDT feature space for the classification task. Finally, we obtain a high discriminative and powerful descriptive HEp-2 cell image representation based on the adaptive CoDT feature space using FK principle. We feed the image representation into a linear SVM classifier to predict staining patterns of the HEp-2 cells. Our proposed AdaCoDT method combines the strengths of generative and discriminative approaches, therefore it can achieve excellent classification performance. Experimental results on the ICPR2012 dataset and the ICIP2013 training dataset validate that the AdaCoDT method can provide superior performance for HEp-2 cells classification, compared with the traditional LBP and its extensions. The new feature encoding method also improves the classification performance in comparison of the BoW representation.