Abstract 7165: AI-driven vascular structure analysis predicts immunotherapy response in non-small cell lung cancer

Abstract Immune checkpoint inhibitors (ICI) have become the standard treatment for advanced non-small cell lung cancer (NSCLC). PD-L1 immunohistochemistry (IHC), the only approved biomarker for predicting ICI treatment response, shows limited predictive value, particularly in patients with low expression (1-49%). Abnormal tumor vasculature in NSCLC correlates with poor prognosis and hampers treatment efficacy by obstructing drug delivery and suppressing immune cell activation. We developed an AI-based biomarker that predicts immunotherapy response by comprehensively analyzing tumor vascular patterns on contrast-enhanced CT. We developed and trained our model using contrast-enhanced CT data from 1, 051 NSCLC patients, focusing on both tumor-associated vasculature and normal lung vessels. We developed the Vessel Risk Score (VRS) based on the Mahalanobis distance between the feature distributions of tumor vessels and normal vessels in immunotherapy-treated patients. For external validation, we analyzed pre-treatment CT images from 321 patients who received immunotherapy at two independent institutions (Ajou University Hospital: 218 patients, Inha University Hospital: 103 patients). All images were reviewed by radiation oncologists. VRS effectively stratified patient outcomes. The VRS Low group (n=160) showed significantly longer progression-free survival (PFS) compared to the VRS High group (n=161) (mPFS: 4.9 vs 2.8 months, p=0.013). In multivariate analysis, VRS proved to be an independent prognostic factor alongside PD-L1 expression for both PFS (HR=1.63, 95% CI=1.26-2.10, p<0.001) and OS (HR=1.60, 95% CI=1.18-2.17, p=0.003). Notably, in patients with low PD-L1 expression (1-49%), where clinical decision-making is challenging, VRS demonstrated superior predictive performance with AUROC of 0.68 (95% CI: 0.60-0.77) for treatment response (CR, PR, SD) and 0.70 (95% CI: 0.60-0.78) for durable clinical benefit (DCB), while PD-L1 showed predictive power below 0.5. VRS showed no significant correlation with PD-L1 (p=0.354), suggesting its complementary role in assessing vascular abnormality. The combination of VRS and PD-L1 expression enhanced prognostic stratification: PD-L1 High/VRS Low patients showed mPFS of 7.2 months and mOS of 23.9 months, while PD-L1 Low/VRS High patients showed mPFS of 2.3 months and mOS of 9.8 months, demonstrating marked prognostic differences. Our CT-based vascular structure biomarker provides an effective, non-invasive method for predicting immunotherapy response in NSCLC patients. VRS is particularly useful in predicting treatment outcomes for patients with low PD-L1 expression and complements PD-L1 expression data. These findings underscore the significance of tumor vascular patterns in NSCLC immunotherapy, potentially enhancing patient selection and facilitating personalized treatment strategies. Citation Format: JunHyeong Park, Jaesung Heo, Jun Hyeok Lim, Woo Kyung Ryu, Jeong-Seon Ryu, Jeong-Seok Choi, Lucia Kim. AI-driven vascular structure analysis predicts immunotherapy response in non-small cell lung cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 7165.