Noninvasive imaging-based vascular score is associated with immunotherapy outcome in non-small cell lung cancer

Abstract Morphological abnormality in tumor vasculature, a recently recognized mechanism of resistance to immune checkpoint inhibitors (ICIs), remains difficult to quantify due to its heterogeneity. We present the vascular risk score (VRS), a deep learning–based imaging metric that quantifies abnormalities in tumor vasculature on CT scans. We trained a deep learning model to learn representations of vascular morphology. Abnormality was then quantified using Gaussian mixture modeling as the degree to which each patient’s tumor vasculature deviated from the learned distribution of normal morphology. We validated VRS in a cohort of 321 NSCLC patients treated with ICIs. Patients with low VRS showed significantly longer progression-free and overall survival, and lower VRS was observed in patients with disease control compared with progressive disease. Combining VRS with PD-L1 expression provided modest additional discrimination over either marker alone. VRS enables noninvasive, objective quantification of tumor vascular abnormality and may serve as a prognostic imaging marker in NSCLC.