Radiomic phenotypes of tumor angiogenesis compared with PD-L1 in pre-treatment prediction of outcomes across immunotherapy regimens in NSCLC: An external validation study.

8581 Background: Tumor angiogenesis is critical to cancer progression and treatment resistance, as evidenced by the success of therapies targeting both immune activation and neoangiogenesis. Conventional biomarkers like PD-L1 unreliably predict long-term patient outcomes such as overall survival (OS). Quantitative Vessel Tortuosity (QVT) isolates tumor-associated vessels and quantifies abnormal vascular architecture from pre-treatment radiography. We developed novel QVT Phenotypes of chaotic tumor angiogenesis and externally validated their use in the pre-treatment prediction of long term survival across multiple immunotherapy treatment strategies. Methods: QVT Phenotype is a radiomic AI biomarker developed and validated using a real-world dataset of 639 NSCLC patients from 6 institutions. Pre-treatment CT scans of 375 patients from institutions 1-3 were used for phenotype discovery (Dataset A). Deep learning models automatically extracted lung lesions and adjacent vessels. 910 QVT metrics of vascular abnormalities (e.g. curvature, twistedness, and branching) were computed and used to identify intrinsic vascular phenotypes via an unsupervised clustering agnostic. Two validation cohorts from external institutions 4-6 were used to evaluate association with 3-year OS: ICI monotherapy (Mono-ICI) recipients of mixed PD-L1 status (Dataset B, n=172) and Chemo-ICI recipients with PD-L1 TPS<50% (Dataset C, n=90). Results: 38% of patients were QVT High, with twisted and erratic growth patterns on pre-treatment CT scans indicating chaotic angiogenesis. Across validation cohorts (Datasets B+C), QVT-High emerged as a strong marker of poor survival (HR=2.26; p=<1E-5). In Dataset B, QVT-High better predicted poor Mono-ICI outcomes (HR=2.23, p=0.00080) than PD-L1 status (HR=1.99, p=0.032), with a 23.0 month reduction in median OS compared to QVT Low. QVT Phenotype maintained significance within the subset (n=61) of PD-L1 High patients (HR=3.02, p=0.017). In Dataset C, QVT-High stratified Chemo-ICI recipients by OS (HR=2.71, p=0.00060), while PD-L1 status failed to reach significance (HR=1.70, p=0.083). QVT-High Chemo-ICI patients had a 16.3 median OS reduction compared to QVT-Low patients. Conclusions: This validation study establishes QVT Phenotype as a non-invasive biomarker using standard-of-care pretreatment radiographic scans. QVT Phenotype of chaotic tumor angiogenesis is both interpretable and treatment-agnostic. QVT Phenotype predicted survival across multiple immunotherapy regimens in NSCLC, outperforming PD-L1. QVT phenotyping can be used to identify patients unlikely to benefit from existing SOC treatments. Future work will explore using QVT Phenotypes to identify patients who may benefit from escalated therapeutic strategies, including those incorporating anti-angiogenic mechanisms.