Analysis of different color unmixing methods for cMET-PDL1-EGFR multiplex assay.

Abstract

e21080 Background: Multiplex brightfield imaging offers the advantage of simultaneously analyzing multiple biomarkers on a single slide, as opposed to single biomarker labeling on multiple consecutive slides. To accurately analyze multiple biomarkers localized to the same cellular compartment, a set of common lung panel markers were selected as a model, cMET-PDL1-EGFR. One of the most crucial preliminary stages for analyzing this assay is identifying the unique chromogens on individual cells. This is a challenging problem due to the co-localization of membranes from all the three biomarkers. It requires advanced color unmixing for creating equivalent singleplex images (referred as synthetic singleplexes) from each triplex image for each biomarker. Methods: We explored different unsupervised methods for unmixing cMET-PDL1-EGFR images. Initially we explored an unsupervised method based on Lee et al. [1], which formulated the color unmixing problem as a non-negative matrix factorization (NMF) method and proposed a system capable of performing the color decomposition in a fully automated manner, where no reference stain color selection is required. Next, we developed cycle-Generative Adversarial (cycle-GAN) [2] models, for each of the three stains – TAMRA (purple), QM-Dabsyl (yellow) and Green. The models were trained on 6 cases of lung having different intensities of the stains. Results: The NMF and cycle-GAN method was validated on 5400 lung images. On analysis of the results and on review with pathologists, we observed that NMF was able to produce reasonable results on synthetic PDL1 and EGFR, but the synthetic Green images did not replicate the adjacent Green images properly. On the other hand, cycle-GAN was able to generate similar synthetic images for all three stains as demonstrated by the similarity metrics in Table 1. Pathologist review of the cycle-GAN results determined that the difference in percentage scores between pairs of adjacent and synthetic singleplex images would be within the range 2%-5% only. Conclusions: In this project, we performed detailed analysis of two unsupervised methods for color unmixing on triplex images. On comparison of the methods based on image similarity measures and pathologist review, we conclude that the cycle-GAN method is competent in generating synthetic images.