Generative Adversarial Networks for Stain Normalisation in Histopathology

TL;DR

This chapter explores different techniques which have been used for stain normalisation in digital pathology, with a focus on approaches which utilise generative adversarial networks (GANs).

Abstract

The rapid growth of digital pathology in recent years has provided an ideal opportunity for the development of artificial intelligence-based tools to improve the accuracy and efficiency of clinical diagnoses. One of the significant roadblocks to current research is the high level of visual variability across digital pathology images, causing models to generalise poorly to unseen data. Stain normalisation aims to standardise the visual profile of digital pathology images without changing the structural content of the images. In this chapter, we explore different techniques which have been used for stain normalisation in digital pathology, with a focus on approaches which utilise generative adversarial networks (GANs). Typically, GAN-based methods outperform non-generative approaches but at the cost of much greater computational requirements. However, it is not clear which method is best for stain normalisation in general, with different GAN and non-GAN approaches outperforming each other in different scenarios and according to different performance metrics. This is an ongoing field of study as researchers aim to identify a method which efficiently and effectively normalises pathology images to make AI models more robust and generalisable.

Figures (4)

• Figure 1: Examples of visual variation caused by different scanners from the MIDOG 2021 Challenge training set Aubreville2023, where each tissue sample was processed in the same laboratory following the same protocol, and then digitised with one of four available scanners. Image adapted from Breen2022.
• Figure 2: Simplified VAE-GAN architecture, with the VAE decoder used as the GAN generator. Figure adapted from Larsen2016.
• Figure 3: Unsupervised image-to-image translation methods for images $x$,$y$ from domains $\mathcal{X}$,$\mathcal{Y}$, with encoders $E$, generators $G$, and latent spaces $\mathcal{Z}$. (a) CycleGAN Zhu2017 uses domain-specific latent spaces, (b) UNIT Liu2017 uses a shared latent space, (c) MUNIT Huang2018 and DRIT Lee2018 decompose the latent space into domain-specific attribute (style) spaces and a single shared content (structure) latent space. Diagram adapted from Lee2020.
• Figure 4: CycleGAN stain normalisation applied to breast cancer tissue, with a different scanner used in the original and target domains. Figure adapted from Breen2022.