Deep cross residual network for HEp-2 cell staining pattern classification

Abstract

Many computer-aided systems have been developed for Human epithelial type 2 (HEp-2) cell classification recently, but there is still a big performance gap between them and specialist doctors. Inspired by the recent successes of convolutional neural network, we proposed a deep cross residual network (DCRNet) for HEp-2 cell classification. A cross connection based residual block was proposed to increase the information flow among different network layers. We used two benchmark datasets to evaluate our system. The state-of-art results, i.e. the average class accuracy of 80.8% in the International Conference on Pattern Recognition (ICPR) 2012 dataset and the mean class accuracy of 85.1% in the Indirect Immunofluorescence Image (I3A) dataset, were achieved. Our result on the ICPR 2012 dataset is so far the best among all works reported in the literature. Our algorithm was winner of the most recent ICPR 2016 contest and the accuracy beat all of the top performers in the previous International Conference on Image Processing (ICIP) 2013 and the ICPR 2014 contests.