AI-powered assessment of tertiary lymphoid structures (TLS) from H&E whole-slide images as a prognostic tool in HNSCC.

Abstract

6024 Background: Tertiary lymphoid structures (TLS) are recognized prognostic markers in Head and Neck Squamous Cell Carcinoma (HNSCC), yet manual assessment from H&E whole-slide images (WSI) remains subjective and labor-intensive, limiting clinical utility. This highlights the need for a reproducible and objective AI-based approach to TLS assessment using routine H&E slides. Methods: We conducted an integrative analysis by combining transcriptomic and histopathologic data. TCGA HNSCC mRNA expression data (n = 566) were analyzed using xCell to compute enrichment scores for B cells, T cells (CD4+ and CD8+), and dendritic cells. A TLS enrichment score was calculated by averaging the z-standardized aggregate scores of these lineages. Patients in the top and bottom quartiles were labeled ‘TLS enriched’ and ‘non-enriched,’ respectively; these labels were used to train a foundation model-based AI using Imagene’s OI Suite powered by CanvOI with a 3:1 train–test split. Survival analyses were performed at the patient level in 443 evaluable patients with high-quality H&E whole-slide images and definitive AI-predicted TLS enrichment status. Univariable and multivariable Cox regression evaluated AI-predicted TLS enrichment as an independent predictor of overall survival. Results: The AI model demonstrated robust performance for TLS assessment, achieving an AUC of 0.77 in the training set (n = 332, 75%) and AUC of 0.85 in the test set (n = 111, 25%). Kaplan–Meier analysis showed that among 443 patients, the AI-predicted TLS-enriched (TLS+) group (n = 146, 33%) demonstrated improved overall survival compared with the TLS-non-enriched (TLS-) group (n = 297, 67%), with median OS 57.9 vs 35.4 months (HR 0.72; 95% CI 0.53–0.98; log-rank P = 0.039). After adjustment for age, sex, and stage, AI-predicted TLS enrichment remained independently associated with improved overall survival (HR, 0.73; 95% CI, 0.54–1.00; P = 0.0499). These findings suggest that the AI model successfully translates molecular TLS signatures into histological predictors, capturing critical tumor microenvironment (TME) features that refine risk stratification beyond standard clinicopathologic factors. Conclusions: AI-based H&E WSI analysis helps identify TLS enrichment as a potential predictor of overall survival in HNSCC. This unbiased computational approach provides a reproducible and objective methodology using H&E slides alone, without the need for additional molecular or immunohistochemical assays, thereby supporting immune risk stratification for precision immunotherapy, particularly in resource-limited settings.