Predicting Stereotactic Body Radiation Therapy Response Using an AI-Based Tumor Vessel Biomarker

Abstract

IntroductionAbnormal tumor vasculature impairs oxygen delivery and induces hypoxia, contributing to treatment resistance and poor prognosis in non-small cell lung cancer (NSCLC). Although radiation therapy can modulate tumor vessels, its effects vary widely due to vascular heterogeneity. Therefore, a reliable and noninvasive method to quantify vascular abnormality is needed to better predict treatment outcomes.MethodsWe developed a deep learning-based imaging biomarker, the Vessel Risk Score (VRS), to quantify tumor vascular abnormality from contrast-enhanced CT scans. Trained on multi-institutional data from 126 NSCLC patients treated with hypofractionated radiotherapy, the model learned vascular morphology patterns from tumor-vessel images. Using these learned patterns, vascular heterogeneity was quantified as the distributional difference from normal vessel morphology. The generalizability of VRS was then evaluated in an external cohort of 128 early-stage NSCLC patients who underwent stereotactic body radiotherapy (SBRT).ResultsVRS showed significantly better prediction of SBRT radiation therapy response compared to vessel density. The VRS of the responder group was 0.494 (95% CI: 0.47-0.52), significantly lower than the non-responder group's 0.578 (95% CI: 0.54-0.62). Additionally, patients with high VRS showed significantly shorter PFS compared to those with low VRS (p < 0.05). In Cox multivariate analysis, VRS emerged as the only significant predictor among vessel density and other clinical variables (p < 0.05).ConclusionThe proposed AI-derived VRS provides a noninvasive and reproducible measure of tumor vascular abnormality, offering improved prediction of radiation therapy response and prognosis compared with vessel density. This approach may extend to prognostic assessment in other cancer types where vascular morphology plays a critical role.

Keywords: SBRT; artificial intelligence; biomarker; non-small cell lung cancer (NSCLC); tumor vessel.

Figure 1.

Flow diagram of patient selection and cohort distribution. From five Korean institutions, a total of 10,534 lung cancer patients were included, and 3151 NSCLC patients receiving radiation therapy were selected for analysis. These were divided into training (n = 126), external validation (n = 128), and normal vessel (n = 1586) cohorts.

Figure 2.

Overview of the deep learning framework for vessel feature learning and vessel risk score calculation. Step 1 shows the vessel feature learning process using tumor vessel images, and Step 2 illustrates the VRS calculation using Mahalanobis distance.

Figure 3.

Predictive performance of VRS versus vessel density for treatment response. (A) Box plots comparing VRS and vessel density between responder and non-responder groups. (B) ROC curve analysis of VRS and vessel density for predicting treatment non-response.

Figure 4.

Kaplan-Meier survival curves comparing VRS and vessel density for PFS and OS. (A) PFS analysis by VRS and vessel density groups. (B) OS analysis by VRS and vessel density groups.