e13670 Background: Tertiary lymphoid structures (TLS) are important biomarkers reflecting the tumor immune microenvironment and are closely associated with patient prognosis. However, current TLS structural assessment often relies on discrete morphologic grading (e.g., maturity-based classification), which fails to capture the continuous spectrum and complexity of TLS architecture. Therefore, a more systematic quantitative framework is needed to characterize TLS structure along a continuum. In this study, we developed a deep learning model for automated TLS detection on routine H&E whole-slide images (WSIs) and constructed a TLS Structural Prognostic Index (TLS-SPI) for precise risk stratification of overall survival (OS). Methods: This study included WSI data from 2,170 patients in the TCGA pan-cancer cohort, covering STAD, COAD, LUAD, LUSC, and BRCA (TNBC subtype). All WSIs were randomly split into training, validation, and test sets at a 7:2:1 ratio, and then tiled into 512×512 patches at 5× magnification for model training and validation. We trained an automated TLS detection model based on an improved YOLO-v12 architecture, and evaluated performance using TLS-level recall, precision, and F1 score. Furthermore, we constructed TLS-SPI using patient-level quantitative TLS features extracted from automated detection. Kaplan–Meier survival analysis and Cox proportional hazards models were used to assess the association between TLS-SPI and OS, and its prognostic stratification ability was validated in the STAD, LUSC, and COAD cohorts. Results: The improved YOLO-v12 model demonstrated strong robustness for automated TLS detection across multiple cancer types, achieving an overall performance of recall = 0.844, precision = 0.902, and F1 = 0.872. The TLS Structural Prognostic Index (TLS-SPI) derived from this model effectively stratified patients into high- and low-risk groups in the TCGA-STAD, TCGA-LUSC, and COAD cohorts. Multivariable analyses further confirmed TLS-SPI as an independent prognostic factor for overall survival (OS) in these cancer types (LUSC cohort: HR = 0.63, 95% CI 0.47–0.84, p < 0.05), highlighting its significant clinical predictive potential. Conclusions: We established a scalable TLS auto-detection framework applicable across multiple TCGA cancer cohorts and proposed a TLS Structural Prognostic Index (TLS-SPI) that predicts OS and stratifies risk in the STAD, LUSC, and COAD cohorts. This approach provides an extensible path toward standardized TLS quantification from routine H&E slides and supports translational prognostic applications.
Jiang et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: