CC-GAN: A Robust Transfer-Learning Framework for HEp-2 Specimen Image Segmentation

Abstract

Human epithelial type 2 (HEp-2) cell images play an important role for the detection of antinuclear autoantibodies in autoimmune diseases. As the HEp-2 cell has hundreds of different patterns, none of currently available HEp-2 datasets contain all of the types. Therefore, existing automatic processing systems for HEp-2 cells, e.g., cell segmentation and classification, needs to be transferred between different data sets. However, the performances of transferred system often dramatically decrease, especially when transferring supervised-approaches, e.g., deep learning network, from large dataset to the small but similar ones. In this paper, a novel transfer-learning framework using generative adversarial networks (cC-GAN) is proposed for robust segmentation of different HEp-2 datasets. The proposed cC-GAN tries to solve the overfitting problem of most deep learning networks and improves their transfer-capacity. An improved U-net, so-called Residual U-net (RU-net), is developed to work as the generator for cC-GAN model. The cC-GAN was first trained and tested using I3A dataset and then directly evaluated using MIVIA dataset, which is much smaller than I3A. The segmentation result demonstrates the excellent transferring-capacity of our cC-GAN framework, i.e., a new state-of-the-art segmentation accuracy of 75.27% was achieved on MIVIA without finetuning.