Histopathology is the basis of clinical diagnosis, especially in cancer treatment. However, traditional chemical staining is often time-consuming, labor-intensive, and can consume valuable tissue samples. A research team at the Hong Kong University of Science and Technology (HKUST) Faculty of Engineering has developed a novel generative AI (GenAI) framework that can generate high-fidelity virtual staining images even when the positions of training image pairs are imperfect, paving the way for faster and tissue-saving histopathology workflows. A study titled “Generative AI for misregistration-tolerant virtual staining to accelerate histopathology workflows” was recently published in an international journal. nature communications.
This research is led by Professor CHEN Hao, Assistant Professor, Department of Computer Science and Engineering, Director of the Medical and Engineering Innovation Collaboration Center and SmartX Labcollaboration with Professor Terence Wong, Associate Dean and Associate Professor, Department of Chemical and Biological Engineering, Deputy Director of the Center for Medical and Engineering Innovation Collaborationalong with researchers from Southern Medical University in Guangzhou, the Chinese University of Hong Kong, and other collaborative partners.
In routine pathology, tissue samples are commonly processed with chemical stains such as hematoxylin and eosin (H&E) to reveal cell nuclei and tissue structure. On the other hand, techniques that combine two special stains, such as periodic acid-Schiff-Arcian blue (PAS-AB), can further highlight specific biological components. Although these steps are important for disease diagnosis and biomedical research, preparing multiple stained sections is typically time-consuming and can also consume limited and valuable biopsy tissue samples.
Virtual staining offers a promising alternative to improve traditional pathology workflows. By applying AI, researchers can digitally convert unlabeled or routinely stained images into targeted stained images. For example, virtual staining can generate H&E-like images from autofluorescence images, convert H&E images to specialized stains like PAS-AB, and generate multiplex immunohistochemistry images. This approach may reduce the need for repeated chemical staining, preserve tissue samples, and provide an additional “virtual channel” for diagnosis, research analysis, and multimodal modeling.
However, increasing the reliability of virtual staining requires addressing the often underestimated assumption that the input image and its target-stained counterpart are truly aligned. Much of the previous work relies on pairs of registered images and assumes that those pairs are accurate enough for pixel-level alignment. However, in pathology, this assumption is rarely realistic. Tissue sections are non-rigid. Sectioning, staining, scanning, slide mounting, tissue folding, and local damage can all cause spatial deformations. If the image pairs are not perfectly aligned, correctly generated nuclei may be unfairly penalized simply because the corresponding nuclei in the target image are slightly misaligned. For pathology AI, this is not a small error since the value of virtual staining depends on reliable cellular architecture, gland borders, immune cell localization, and spatial distribution of the staining signal.
To address this problem, the research team proposed decoupled generation and registration (DGR). Rather than assuming that the training image pairs are perfectly aligned, DGR explicitly takes into account the registration errors that remain during model training. This framework decouples image generation from spatial alignment. The generative model focuses on learning the appearance and signal transformation between different stains, and the alignment mechanism handles spatial deviations caused by tissue deformation.
The team evaluated DGR across five datasets and four staining-related tasks, including virtual H&E staining from label-free autofluorescence images, conversion of special stains from H&E to PAS-AB, conversion of H&E to multiplexed immunohistochemistry, and normalization of H&E staining. Compared to state-of-the-art virtual staining models, DGR demonstrated superior overall performance in image quality and structural fidelity across multiple tasks.
To further assess visual quality, the researchers asked an experienced pathologist to perform a blind assessment. Pathologists randomly evaluated 500 H&E-stained images and 500 PAS-AB-stained images by comparing the DGR-generated virtual staining with the corresponding real chemical staining. The accuracy of distinguishing between virtual and chemical stains was approximately 52%, close to chance, suggesting that the two are difficult to visually distinguish in this evaluation.
The team also investigated the value of DGR-generated virtual staining in downstream pathology AI analysis. Combining DGR-generated virtual multiplexed immunohistochemistry images with H&E images improved model performance for colorectal polyp classification and gastric cancer tissue classification tasks. These results suggest that the virtual staining generated by DGR is not only visually similar to the real staining, but also preserves morphological and spatial information useful for downstream analyses.
Professor Chen HaoThe paper’s corresponding author said, “This study addresses a key barrier to bringing virtual staining closer to real-world clinical workflows. By enabling high-quality virtual staining from imperfectly aligned pathology images, GenAI provides a more scalable path toward faster and more cost-effective pathology diagnosis.”
reference: Ma J, Li W, Li J et al. Generated AI realizes virtual staining that is resistant to misalignment and accelerates histopathological workflow. Nat Commune. 2026;17(1):4494. doi: 10.1038/s41467-026-71038-2
This article has been reprinted from the following material: Note: Materials may be edited for length and content. Please contact the citation source for details. You can access our press release publishing policy here.
