A Hybrid Ensemble Learning with Generative Adversarial Networks for HEp-2 Cell Image Classification

The Indirect Immunofluorescence (IIF) on Human Epithelial (HEp-2) cells is considered the hallmark protocol of the Anti-Nuclear Antibodies (ANAs) testing for diagnosing autoimmune diseases. The usual practice of visual slide inspection under the fluorescence microscope suffers from low throughput and high labor-subjectivity. Therefore, developing an efficient framework for automatic HEp-2 cell image classification is necessary for overcoming such manual protocol shortcomings. In this paper, a novel HEp-2 cell image classification framework is proposed based on ensemble deep learning with generative adversarial networks (GANs). An efficient Info-WGANGP approach is adopted for data augmentation by generating new HEp-2 cell images and enlarging the size of the training set. Meanwhile, an ensemble deep learning strategy is implemented to build a backbone network obtaining a potent combination of the deep features using three well-known deep convolutional networks, i.e., DCRNet, DSRNet, and HEpNet. The evaluation experiments on the publicly available I3A dataset demonstrate promising classification results in terms of average classification accuracy (ACA) with 98.82% and mean class accuracy (MCA) with 98.91% outperforming the latest deep learning approaches. The proposed classification framework seems to be applicable for supporting human experts in making accurate and rapid diagnosis decisions of the HEp-2 cell patterns.