Imaging-Based Artificial Intelligence in Vascular and Interventional Radiology: A Narrative Review

Artificial intelligence (AI) has shown remarkable success in diagnostic radiology through advanced pattern recognition capabilities, yet its application in vascular interventional radiology (VIR) remains limited due to smaller, more variable datasets. This review examines AI applications in VIR procedures that utilize imaging modalities as input across preprocedural, intraprocedural, and post-procedural stages. A comprehensive literature search across PubMed, Embase, and Web of Science identified studies employing AI models with direct patient care impact, categorized by imaging modality (CT, MRI, fluoroscopy/DSA, ultrasound, X-ray, and multimodal) and task type (segmentation, detection, prediction, and miscellaneous). AI demonstrated substantial promise across multiple VIR domains. Deep learning models achieved high Dice similarity coefficients (0.82-0.962) for anatomical structure segmentation including aortic dissections and abdominal aortic aneurysms. Detection tasks showed excellent performance with accuracies up to 95% for endoleak detection and AUCs reaching 0.97 for vessel stenosis identification. Prediction models frequently outperformed traditional clinical assessments, with AUCs exceeding 0.90 for outcomes after EVAR, TEVAR, TACE, and TARE procedures. Radiomics-based approaches combined with machine learning proved particularly effective for treatment response prediction and risk stratification. Despite challenges including limited dataset sizes, potential bias, and interpretability concerns, AI shows transformative potential in VIR. Continued clinician AI expert collaboration will be crucial for responsible deployment and optimization of patient care in interventional radiology.