Quantifying Localized Tumor-Infiltrating Lymphocytes by Deep Learning on H&E Images: Analytical Validation and Prognostic Value in Lung Adenocarcinoma

Tumor-infiltrating lymphocytes (TILs) are a critical determinant of response to immune checkpoint inhibitor (ICI) therapy in lung adenocarcinoma (LUAD). However, assessing the clinical utility of TIL quantification is challenging due to subjective manual scoring and differences in methodology, including some that may obscure localized immune responses critical to ICI response. Together, these issues limit our understanding of potential prognostic benefits of performing routine TIL quantification. Here, we present the "focal TIL score", an artificial intelligence-powered biomarker derived from routinely generated Hematoxylin & Eosin (H&E) pathology images which measures TIL density within tiles across the entire relevant tissue region in order to account for spatial heterogeneity of localized lymphocytic infiltration. Underlying this method are two deep learning models that ensure accurate, reproducible, and interpretable quantification: one for lymphocyte detection trained on IHC-derived ground truth, and one for segmentation of tumor and tumor-adjacent stroma region ("scorable region") trained on consensus pathologist annotations. We performed an analytical validation of the two models, demonstrating high overall performance: lymphocyte model concordance correlation coefficient (CCC) = 0.90, scorable region model Dice score = 0.87. We then performed a prognostic association study in a cohort of 685 ICI-treated LUAD patients and saw that, the focal TIL score was significantly associated with overall survival (HR=0.60, 95% CI 0.48-0.75, P<0.001), outperforming the standard biomarker PD-L1. Our findings show that quantifying the spatial intensity of immune infiltration, rather than the global average, provides a superior measure of the anti-tumor immune response. The accessibility and standardization associated with this approach may enable further studies and developments in precision oncology for LUAD.