Article Text

Download PDFPDF

1299 Predicting CD8+ cell density and tumor-immune phenotypes in non-small-cell lung cancer (NSCLC) from standard H&E slides using deep learning (DL)
  1. David Soong1,
  2. Becky Arbiv2,
  3. Ettai Markovits2,
  4. Alon Groisman2,
  5. Yuval Shachaf2,
  6. Tomer Dicker2,
  7. Yoni Yedidia2,
  8. Hisham Hamadeh1,
  9. Kate Sasser1,
  10. Gali Golan2,
  11. Brandon Higgs1,
  12. Suzana Couto1 and
  13. Ori Zelichov2
  1. 1Genmab, Princeton, NJ, USA
  2. 2Nucleai, Tel Aviv, Israel


Background Tumor infiltrated lymphocytes (TIL), namely CD8+ TILs play a major role in antitumor immunity and tumor cell eradication. High-density infiltration of CD8+ cells in the tumor, in contrast to CD8+ cell excluded regions, is associated with improved prognosis and response to immunotherapy in multiple cancer types, however, CD8 evaluations require IHC staining, often not performed routinely in clinical practice. Here, we used DL to predict CD8+ cell density and immune phenotypes from standard H&E slides.

Methods 188 pairs of H&E slide and a sequential CD8 stained slide from 103 patients with metastatic NSCLC were procured. DL models were trained to classify tumor cells, lymphocytes, fibroblasts, and tumor versus stromal areas on H&E, as well as positivity of CD8 per cell by IHC. 354 spatial features were calculated from the H&E slides and CD8+ density in the whole tumor region from IHC slides. A training (n=143) and test (n=45) cohort was created and linear regression modeling predicted CD8 density from H&E features. Two board certified pathologists classified IHC slides into immune phenotypes: inflamed, desert and excluded, based on CD8 density (table 1) and a multinomial logistic regression model was train to predicet these phenotypes from the H&E images.

Results The H&E features most predictive of CD8 density were related to lymphocyte densities, tumor-lymphocyte proximity in the invasive margin, and lymphocyte density in the tumor area. Correlation using this 3 H&E feature model on the test set was r=0.87 (p<0.0001, figure 1). The H&E features most predictive to immune phenotype were proximity between lymphocytes and tumor cells or granulocytes, as well as lymphocyte:fibroblast density ratio. Model accuracy on the test set reached an accuracy of 74% (95% CI 57%-85%, p<0.0001) in classifying slides to the correct immune phenotypes.

Conclusions Spatial analysis of immune and tumor cells from standard H&E slides using DL can accurately predict CD8 cell density and identify immune phenotypes. By only requiring H&E images, these biomarkers important for checkpoint inhibitor therapy can be measured in clinical practice without the requirement of IHC staining.

Abstract 1299 Table 1

Criteria for Immune Phenotyping

Abstract 1299 Figure 1

Correlation between test set predicted CD8 density from H&E analysis and real test set CD8 density calculated on IHC slides

Statistics from

Request Permissions

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.