Background Multiplex immunofluorescence (mIF) is a powerful tool for tumor-microenvironment (TME) characterization. However, the adoption of high-plex panels in the clinical setting is limited due to cost and technical complexity. Hematoxylin and eosin (H&E) staining provides morphologic information about tissue structure and cellular composition but despite its use in routine pathology, its potential for extracting meaningful data in combination with mIF has been largely overlooked. We leveraged same-slide H&E and mIF images to enrich our analysis, enabling identification of known biomarkers, unidentifiable from low-plex mIF alone.
Methods We stained consecutive slides from 57 melanoma biopsies using two 6-plex immunofluorescence antibody panels and same-slide H&E staining. The panels consisted of markers for myeloid and tumor cells, with the first panel including T-cell markers (CD4, CD8) and the second panel composed of B-cell markers (CD20, CD38). Using a deep learning-based multiplex imaging analysis pipeline, we generated predictions for 6 cell types for each mIF panel. Simultaneously, deep learning models processed the H&E slides to detect and classify cells. The mIF and H&E slides were registered, and mIF marker-negative cells were reclassified based on their registered H&E cell type. Tertiary lymphoid structures (TLS) were annotated using markers from both panels and predicted using a model trained on T-cell panel data enhanced with cells from H&E.
Results Out of the 20 million cells detected in the B-cell panel, 9 million were successfully classified as tumor-, dendritic-, plasma-, or B-cells. However, a significant number of cells were classified as marker-negative cells. Integrating H&E cell typing, we reclassified 6 million cells as SOX10-negative tumor cells, fibroblasts, granulocytes, and non-B-lymphocytes. The non-B-lymphocytes, derived from H&E images, showed strong correlation with T-cell numbers identified by the T-cell mIF panel, enabling us to assess T-cell infiltration in the tumor invasive margin using the B-cell panel (figures 1 and 2). In the T-cell panel, we reclassified 5.5 million cells, including non-T-lymphocytes, which showed high correlation with B-cell numbers in the B-cell panel (figure 2). Leveraging these reclassified cells, we developed a TLS identification model, demonstrating 88% accuracy compared to mIF-based manual annotations (figure 3).
Conclusions Our analysis demonstrates the value of integrating same-slide H&E analysis in low-plex mIF studies. By leveraging the morphologic information provided by H&E slides, we were able to identify additional cells and accurately quantify known biomarkers for immunotherapy response, such as TLS and T-cell tumor infiltration, which were not identifiable otherwise.
Markovits E, Dankovich T, Gluskin R, et al. A novel deep learning pipeline for cell typing and phenotypic marker quantification in multiplex imaging. bioRxiv doi:10.1101/2022.11.09.515776
This is an open access article distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited, appropriate credit is given, any changes made indicated, and the use is non-commercial. See http://creativecommons.org/licenses/by-nc/4.0/.
Statistics from Altmetric.com
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.