HEp-2 Cell Image Classification with Convolutional Neural Networks

Abstract

The diagnosis of many autoimmune diseases can be greatly facilitated by automatic staining patterns classification of Human Epithelial-2 (HEp-2) cells within indirect immunofluorescence (IIF) images. In this paper, we propose a framework to classify the HEp-2 cells by utilizing the deep convolutional neural networks (CNNs). With carefully designed network architecture and optimized parameters, our networks extract features from raw pixels of cell images in a hierarchical manner and perform classification jointly, avoiding using hand-crafted features to represent a HEp-2 cell image. We evaluate our method on the training dataset of HEp-2 cells classification competition held by ICPR 2014. Our system achieves mean class accuracy of 96.7% on the held-out test set and it also obtains competitive performance on the ICPR 2012 cell dataset.