利用mRNA衍生标签从H&E全切片图像中无偏地AI检测三级淋巴结构可预测NSCLC生存

Tertiary lymphoid structures (TLS) are recognized prognostic markers in non-small cell lung cancer (NSCLC), yet manual detection from H&E whole-slide images (WSI) remains subjective and labor-intensive, limiting clinical adoption. We conducted an integrative analysis combining transcriptomic and histopathologic approaches. TCGA NSCLC mRNA expression data (n = 922) were analyzed using xCell to compute enrichment scores for B cells, T cells (CD4+ and CD8+), and dendritic cells. TLS enrichment score was calculated by averaging z-standardized xCell aggregate enrichment scores for B cells, T cells, and dendritic cells. Samples in the upper quartile were labeled ‘TLS enriched' and those in the lower quartile were labeled ‘TLS non-enriched.' These quartile-derived labels were then used to train a machine learning model for TLS detection from TCGA H&E whole-slide images using Imagene's OI Suite platform with a 3:1 train-test split. Univariable and multivariable cox proportional hazards regression analysis evaluated whether AI-predicted TLS enrichment independently predicted overall survival (OS). The AI model demonstrated robust performance for TLS detection, achieving an AUC of 0.84 in the training set and a test AUC of 0.92. Kaplan-Meier analysis showed that patients in the AI-predicted TLS-enriched group (n = 366, 366/922=39.7%) demonstrated improved OS compared with the AI-predicted TLS-non-enriched group (n = 556, 556/922=60.3%; HR 0.76; 95% CI 0.61-0.94; log-rank p = 0.014). Adjusting for age, sex, and histology, AI-predicted TLS enrichment remained independently associated with favorable OS (HR 0.77; 95% CI, 0.61-0.97; p = 0.025). Among the covariates, older age (greater than the cohort median) was associated with worse survival (HR 1.25; 95% CI, 1.01-1.55; p = 0.039), while sex and histologic subtype (LUSC vs LUAD) were not significant predictors (HR 1.12; 95% CI, 0.89-1.41; p = 0.337 and HR 0.99; 95% CI, 0.79-1.25; p = 0.957, respectively). These findings indicate that the model's TLS-enrichment prediction captures clinically meaningful tumor microenvironment features that stratify OS beyond standard clinicopathologic factors. AI-assisted H&E WSI analysis helps identify TLS enrichment as a potential independent predictor of overall survival in NSCLC. This unbiased computational approach provides a reproducible and objective methodology for TLS assessment and survival prediction in a resource-limited context using H&E WSI only, and may support downstream treatment selection in the context of precision immunotherapy.