A Novel Generative Artificial Intelligence Method for Interference Study on Multiplex Brightfield Immunohistochemistry Images

Overview

• Multiplex brightfield imaging allows analyzing multiple biomarkers on a single slide.
• Accurately analyzing co-localized biomarkers is challenging due to color overlaps.
• This project developed a cycle-Generative Adversarial Network (cycle-GAN) method to unmix triplex images and generate singleplex images for each biomarker.
• The method uses optical density domain inputs to improve image quality compared to RGB inputs.
• The cycle-GAN models were validated on over 14,000 lung, gastric, and colon tissue images.

Plain English Explanation

Biomarkers are molecules that provide information about a person's health or disease. Multiplex imaging allows examining multiple biomarkers at once on a single tissue sample, rather than looking at them one by one. This is useful for getting a more complete picture.

However, when multiple biomarkers are present in the same cell, their colors can overlap and make it hard to tell them apart. The researchers developed a machine learning method to digitally separate the colors and generate individual images for each biomarker.

This method uses a type of AI model called a cycle-GAN. The key innovation is that it takes in the images in a different format (optical density) that helps produce clearer separated images. The model was tested on over 14,000 tissue samples and proved effective at accurately unmixing the biomarker colors.

Technical Explanation

The researchers selected two representative biomarker sets - cMET-PDL1-EGFR and CD8-LAG3-PDL1 - where the three biomarkers can co-localize on the cell membrane. Accurately analyzing co-localized biomarkers requires advanced color unmixing to create equivalent singleplex images from each triplex image.

The researchers developed a cycle-GAN method to unmix the triplex images. Three different models were designed to generate the singleplex image for each of the three stains - Tamra (purple), QM-Dabsyl (yellow), and Green.

A key novelty was using optical density domain inputs instead of RGB images. This helped reduce blurriness in the synthetic singleplex images compared to training on RGB.

The cycle-GAN models were validated on 10,800 lung, gastric and colon images for the cMET-PDL1-EGFR assay and 3600 colon images for the CD8-LAG3-PDL1 assay. Visual and quantitative assessments showed the method is effective and efficient compared to manual review, making it applicable to various multiplex assays.

Critical Analysis

The paper provides a robust technical validation of the cycle-GAN method for biomarker color unmixing. However, some potential limitations or areas for further research include:

• The method was only tested on a limited set of biomarker combinations. Its generalizability to a wider range of multiplex assays could be further explored.
• The quality and consistency of the synthetic singleplex images generated by the cycle-GAN models were assessed, but their impact on downstream biomarker quantification and analysis was not directly evaluated.
• While optical density inputs improved image quality, the need for pre-processing images in this format may limit the practical deployment of the method in some settings.

Overall, the research represents a meaningful advance in multiplex image analysis, but continued refinement and real-world testing would help further demonstrate its capabilities and limitations.

Conclusion

This study developed an effective cycle-GAN-based method to unmix multiplex brightfield images and generate high-quality singleplex images for individual biomarker analysis. The use of optical density domain inputs was a key innovation that improved the quality of the synthetic images.

Validating the method on a large dataset of tissue samples showed it can be reliably applied to various multiplex assays, streamlining the analysis of co-localized biomarkers. This advance has the potential to enhance our understanding of complex biological systems and accelerate biomarker-based diagnostics and disease monitoring.