Research Article
The Generation of Virtual Immunohistochemical Staining Images Based on an Improved Cycle-GAN
@INPROCEEDINGS{10.1007/978-3-030-66785-6_16, author={Shuting Liu and Xi Li and Aiping Zheng and Fan Yang and Yiqing Liu and Tian Guan and Yonghong He}, title={The Generation of Virtual Immunohistochemical Staining Images Based on an Improved Cycle-GAN}, proceedings={Machine Learning and Intelligent Communications. 5th International Conference, MLICOM 2020, Shenzhen, China, September 26-27, 2020, Proceedings}, proceedings_a={MLICOM}, year={2021}, month={1}, keywords={Virtual staining Immunohistochemistry Cycle-GAN}, doi={10.1007/978-3-030-66785-6_16} }- Shuting Liu
Xi Li
Aiping Zheng
Fan Yang
Yiqing Liu
Tian Guan
Yonghong He
Year: 2021
The Generation of Virtual Immunohistochemical Staining Images Based on an Improved Cycle-GAN
MLICOM
Springer
DOI: 10.1007/978-3-030-66785-6_16
Abstract
Pathological examination is the gold standard for the diagnosis of cancer. In general, common pathological examinations include hematoxylin-eosin (H&E) staining and immunohistochemistry. H&E staining examination has the advantages of short dyeing duration and low cost, which is the most common one in the clinical practice. However, in some cases, the pathologist is hard to conduct an accurate diagnosis of cancer only according to the H&E staining images. Whereas, the immunohistochemistry examination can further provide enough evidence for the diagnosis process. Hence, the generation of virtual Ki-67 staining sections from H&E staining sections by computer assisted technology will be a good creative solution. Currently, this is still a challenge due to the lack of pixel-level paired data. In this paper, we propose a new method based on Cycle-GAN to generate Ki-67 staining images from the available H&E images, and our method is validated on a neuroendocrine tumor dataset. Massive experiment results show that the addition of skip connection and structural consistency constraint can further improve the performance of Cycle-GAN in unpaired pathological image-to-image transfer tasks. The quantification evaluation demonstrates that our proposed method achieves the state of art and reveals significant potential in clinical virtual staining.
