AI-based non-invasive profiling of the tumor immune microenvironment using longitudinal CT radiomics predicts immunotherapy response in lung cancer

Abstract

Background: Despite advances in immunotherapy, durable responses in lung cancer remain limited to a subset of patients, underscoring the need for biomarkers capturing spatial immune-tumor interactions. Current methods, such as PD-L1 immunohistochemistry, suffer from sampling bias and fail to decode dynamic immune evasion mechanisms non-invasively. Methods: We developed a radiomics framework integrating longitudinal tumor growth kinetics (log volume change rate, LVCR) with deep learning to: (1) delineate tumors via medical knowledge-guided segmentation; and (2) derive an Immune Evasion Score (IES) predicting immunosuppressive niches. The model employs immune-aware attention gates (IAAG) to prioritize regions associated with aggressive growth (high LVCR) and immune evasion. Results: Validated on 420 CT scans, our approach achieved superior segmentation accuracy (Dice=0.7728 ± 0.03; HD95 = 9.8 ± 1.5 mm) over existing models. Critically, the IES predicted PD-L1 expression (AUC = 0.85; *p*<0.001) and CD8+ T-cell exclusion (*p*<0.01). High IES correlated with rapid immunotherapy progression (HR = 2.3, *p*=0.004), and spatial analysis confirmed 72.3% concordance between IAAG-prioritized regions and pathological PD-L1+ niches. Conclusion: This work establishes a non-invasive paradigm for mapping immunosuppressive microenvironments, bridging precision radiotherapy with immunotherapy personalization. The IES provides a dynamic biomarker of immune evasion, potentially guiding patient stratification for checkpoint inhibitors.