Class-Specific Hierarchical Classification of HEp-2 Cell Images: The Case of Two Classes

We propose and analyze a novel framework for classification of HEp-2 cell images. It is based upon two important aspects. First, we propose to utilize the expert knowledge about the visual characteristics of classes to formulate class-specific image features. Secondly, realizing that the problem involves a small number of classes, we treat the classification task as hierarchical verification subtasks. Thus, the overall classification problem is posed as a verification of each class, using its class-specific features. The current study reports the results using the Nuclear Membrane and Golgi classes. We demonstrate that our framework yields high classification rate with simple and efficient feature definitions, and only (20%) of the data for training. We also analyze important aspects such as comparison with non-hierarchical approach, and performance on low-contrast images which are important for early disease detection.