Development and validation of an interpretable prediction model using spatial patterns of tumor-infiltrating lymphocytes in H&E-stained whole-slide images for immune subtyping of lung adenocarcinoma.

OBJECTIVE: To develop an interpretable prediction model for lung adenocarcinoma immune subtyping by quantifying spatial distribution patterns of tumor-infiltrating lymphocytes in H&E whole-slide images, providing a computational tool for tumor immune microenvironment evaluation.

METHODS: Immune subtyping was performed on the TCGA lung adenocarcinoma cohort using ssGSEA to quantify immune gene set activity, followed by hierarchical clustering and t-SNE visualization to stratify patients into high- and low-immunity subgroups. Immune cell infiltration was assessed using CIBERSORT, while tumor mutation burden and somatic mutation profiles were analyzed with maftools. Differential expression and functional enrichment analyses were conducted using GO and KEGG databases. In pathological image analysis, an automated annotation model optimized with study-specific data was employed to process whole-slide images. Immune subtype prediction criteria were established by quantifying spatial distribution features of tumor-infiltrating lymphocytes. The model's predictive performance was validated in both internal and external cohorts.

RESULTS: Transcriptomic analysis stratified 503 LUAD patients into high- and low-immunity subgroups. The high-immunity group exhibited elevated infiltration of CD8+T cells and M1 macrophages, higher tumor mutation burden, and enriched T cell activation pathways. The low-immunity group showed predominant resting immune cells. The automated annotation model, achieved a 95.09% Dice score for tissue contour segmentation, 91.53% for tumor parenchyma segmentation, and a 79.51% F1-score with an mAP@0.5 of 82.13% for TIL identification. Quantitative TIL spatial distribution analysis with a 0.2 high-attention threshold enabled development of an immune subtyping model using a 0.05 classification cutoff, which achieved an AUC of 0.839 for immune subtype classification in the internal validation cohort. In the external validation cohort, the model achieved an AUC of 0.927, and immunohistochemical analysis confirmed significantly higher densities of CD3+, CD8+, CD20+, and CD68+ cells in predicted high-immunity samples.

CONCLUSION: This study establishes an interpretable immune subtype prediction model for LUAD based on TIL spatial distribution in H&E-stained sections. Through a modular design that integrates deep learning-based annotation with statistical classification, the model links morphological phenotypes to molecular immune subtypes while maintaining transparency and verifiability throughout the analytical workflow. This cost-effective and scalable tool offers potential value for assessing tumor immune status and guiding immunotherapy decision-making.