Deep Adversarial Network Based Stain Unmixing for Brightfield Multiplex Immunohistochemistry Images

Multiplex immunohistochemistry (IHC) makes it possible to simultaneously label multiple protein biomarkers with different colored stains in a tissue section. Unmixing the multiplex IHC image provides an efficient way to obtain the rich diagnostic information each biomarker contains. However, due to the limitation of three-channel RGB images taken by a CCD color camera, it is challenging to unmix brightfield multiplex IHC images with more than three stains. The main technical challenge is that the unmixing is inherent underdeterminate, leading to the possibility of multiple solutions. In this paper, we propose a novel unmixing method using generative adversarial networks (GANs) for brightfield multiplex IHC images containing more than three stains. Our method takes advantage of slides stained only with individual biomarker to address the intriguing task without any human annotation. We propose to employ adversarial training to automatically learn the optimal unmixing, without relying on any inadequately designed priori. To the best of our knowledge, the method achieves state-of-the-art (SOTA) results in terms of unmixing quality, speed, and practicality, as evidenced by both pathologists' visual comparisons and quantitative experiments.