Deep learning-based quantification of tumor-infiltrating lymphocytes as a prognostic indicator in nasopharyngeal carcinoma: multicohort findings

Nasopharyngeal carcinoma (NPC) features a tumor-immune microenvironment rich in tumor-infiltrating lymphocytes (TILs), important for prognosis but labor-intensive to quantify. This study evaluates a deep learning model to quantify TILs (TILDL) in hematoxylin and eosin (H&E)-stained whole-slide images (WSIs) of NPC and explores the association of TILDL percentage with patient outcomes and response to immune checkpoint blockade (ICB). We retrospectively analyzed 435 nonmetastatic NPC patients from two centers, divided into a training cohort (n = 220) and a validation cohort (n = 215). An additional cohort of de novo metastatic NPC patients receiving ICB therapy (n = 63) was included. The deep learning model calculated TILDL percentages from H&E-stained WSIs. Correlations between TILDL percentages and immunohistochemistry (IHC)-derived TIL densities were assessed. Survival analyses evaluated their prognostic significance. TILDL percentages showed strong correlations with IHC-derived TIL densities (CD3+ T cells R = 0.46, CD8+ T cells R = 0.33, CD20+ B cells R = 0.57; all P < 0.001). Higher TILDL percentages (median ≥45.7%) were associated with better 5-year disease-free survival (DFS) and overall survival (OS) in both training (DFS: 80.6% versus 62.5%, P = 0.016; OS: 84.4% versus 71.8%, P = 0.025) and validation cohorts (DFS: 87.3% versus 74.3%, P = 0.016; OS: 93.7% versus 82.6%, P = 0.010). In the ICB-treated metastatic cohort, higher TILDL percentages predicted better 3-year progression-free survival (PFS: 40.5% versus 25.0%, P = 0.022). Multivariate analyses confirmed TILDL percentage as an independent prognostic factor in both settings. The TILDL percentage derived from H&E-stained WSIs effectively stratifies risk in nonmetastatic NPC and may serve as a biomarker in metastatic NPC treated with ICB, aiding in patient selection for individualized treatment.