Comparing virtual staining artificial intelligence model performance on same-tissue versus serial-tissue sections based on CD3+ T-cell ground truth

Abstract

Immunophenotyping via multi-marker assays significantly contributes to patient selection, therapeutic monitoring, biomarker discovery, and personalized treatments. Despite its potential, the multiplex immunofluorescence (mIF) technique faces adoption challenges due to technical and financial constraints. Alternatively, hematoxylin and eosin (H&E)-based prediction models of cell phenotypes can provide crucial insights into tumor-immune cell interactions and advance immunotherapy. Current methods mostly rely on manually annotated cell label ground truths, with limitations including high variability and substantial labor costs. To mitigate these issues, researchers are increasingly turning to digitized cell-level data for accurate in-situ cell type prediction. Typically, immunohistochemical (IHC) staining is applied to a tissue section serial to one stained with H&E. However, this method may introduce distortions and tissue section shifts, challenging the assumption of consistent cellular locations. Conversely, mIF overcomes these limitations by allowing for mIF and H&E staining on the same tissue section. Importantly, the multiplexing capability of mIF allows for a thorough analysis of the tumor microenvironment by quantifying multiple cell markers within the same tissue section. In this study, we introduce a Pix2Pix generative adversarial network (P2P-GAN)-based virtual staining model, using CD3+ T-cells in lung cancer as a proof-of-concept. Using an independent CD3 IHC-stained lung cohort, we demonstrate that the model trained with cell label ground-truth from the same tissue section as H&E staining performed significantly better in both CD3+ and CD3- T-cell prediction. Moreover, the model also displayed prognostic significance on a public lung cohort, demonstrating its potential clinical utility. Notably, our proposed P2P-GAN virtual staining model facilitates image-to-image translation, enabling further spatial analysis of the predicted immune cells, deepening our understanding of tumor-immune interactions, and propelling advancements in personalized immunotherapy. This concept holds potential for the prediction of other cell phenotypes, including CD4+, CD8+, and CD20+ cells.