Comparison of Tumor-Infiltrating Lymphocyte Quantification Using AI-Based H&E Analysis Versus CD8 Immunohistochemistry Across Multiple Tumor Types

Summary: Tumor-infiltrating lymphocyte (TIL) quantification in solid malignancies plays a key prognostic role and is essential for selecting cases for immunotherapy. At our institution, we routinely digitally quantify TILs using clinically validated CD8 immunostaining protocols on multiple tumor types. Recently, open-source AI algorithms have been developed to quantify TILs directly from H&E slides, potentially bypassing the need for CD8 immunostaining. This study aimed to assess the concordance of an AI algorithm with the established CD8 immunostaining protocol. We selected 240 colorectal adenocarcinomas, 59 gastrointestinal stromal tumors, 52 cutaneous melanomas, and 36 head and neck squamous cell carcinomas to evaluate diverse tumor morphologies. H&E staining and CD8 immunostaining (clone 4B11, Leica) were performed on immediately subjacent tumor sections. Tumor and stroma were manually annotated over the same tissue areas of H&E and CD8 slides. H&E-based TIL quantification was conducted using a pre-trained patch-based neural network, available open-access through the WSI-INFER QuPath extension, to classify patches on whole slide images into "lymphocytes" versus "others." The QuPath positive cell detection algorithm was utilized for TIL quantification on digitally scanned CD8-immunostained slides. TIL density was calculated per unit area (mm2) for H&E and CD8-based methods, followed by Spearman correlation testing to evaluate concordance between methods. Correlation strength was defined by the r coefficient: very strong (0.9-1.0), strong (0.7-<0.9), moderate (0.4-<0.7), and weak (0-<0.4). P<0.05 was deemed significant. Strong concordance between methods was observed in head and neck squamous cell carcinomas (r=0.809, p<0.01), while colorectal adenocarcinomas (r=0.448, p<0.01) and gastrointestinal stromal tumors (r=0.504, p<0.01) exhibited moderate correlation. Weak correlation was established between methods in melanomas (r=0.354, p=0.01). Melanoma cases with discrepant results were manually reviewed to identify potential failure modes. Causes of failure included the cytomorphologic similarity of melanoma tumor cells to lymphocytes, leading to misclassification of lymphocytes as "other," and the presence of extensive melanin pigmentation, which erroneously amplified the CD8-based quantification results. AI-based TIL quantification on H&E slides demonstrates variable concordance with CD8 immunostaining, ranging from weak to strong depending on tumor type. While the AI-based approach offers potential cost savings by eliminating the need for ancillary immunohistochemistry, its implementation in clinical settings should be preceded by further validation studies, especially for tumor types like melanoma where performance is inconsistent. Azfar Neyaz, Liron Pantanowitz, Matthew G Hanna, Thomas Pearce, Hooman Henry Rashidi, Lindsey Seigh, Hamdi Surakji, Daniel Christensen, M-Nasan Abdul Baki, Ihsan Baroudi, Samer N Khader, and Ibrahim Abukhiran