Efficient Human Epithelial-2 (HEp-2) cell image classification can facilitatethe diagnosis of many autoimmune diseases. This paper presents an automaticframework for this classification task, by utilizing the deep convolutionalneural networks (CNNs) which have recently attracted intensive attention invisual recognition. This paper elaborates the important components of thisframework, discusses multiple key factors that impact the efficiency oftraining a deep CNN, and systematically compares this framework with thewell-established image classification models in the literature. Experiments onbenchmark datasets show that i) the proposed framework can effectivelyoutperform existing models by properly applying data augmentation; ii) ourCNN-based framework demonstrates excellent adaptability across differentdatasets, which is highly desirable for classification under varying laboratorysettings. Our system is ranked high in the cell image classificationcompetition hosted by ICPR 2014.
展开▼