TRANSFER LEARNING OF A CONVOLUTIONAL NEURAL NETWORK FOR HEP-2 CELL IMAGE CLASSIFICATION

The recognition of the staining patterns of Human Epithelial-2 (HEp-2) cells in indirect immunofluorescence (IIF) images is essential for the diagnosis of several autoimmune diseases. The main challenge is the extraction and selection of the optimal feature set that not only represents the cells' characteristics, but also distinguishes between the classes of cell images with similar appearances. In this paper, we propose a system to classify HEp-2 cell images by applying transfer learning from a pre-trained deep convolutional neural network (CNN) to extract the generic features and then using a feature selection method to get the most relevant features for classification. Although the CNN was trained with a dataset very different from cell images, our system is capable of extracting important semantic features that represent a HEp-2 cell image. When evaluated on the ICPR2012 cell dataset, our method outperforms all other methods on the dataset of the 2012 competition, and demonstrates stable performance under different test protocols.