Abstract
BACKGROUND AND OBJECTIVE: Online adaptive radiotherapy (ART) is an emerging tool to adapt radiotherapy plans to interfraction changes during the treatment course, reducing dose to organs at risk (OARs) while ensuring adequate dose to the tumor volume. Artificial intelligence (AI) can be a promising tool to support online ART techniques given that daily treatment changes can be resource- and time-intensive. With increasing integration of AI into radiation treatments, this literature review summarizes current applications and promising future directions of AI for online ART. METHODS: A comprehensive search of databases (EMBASE, MEDLINE, PsycInfo, PubMed, CINAHL, CENTRAL, Scopus, and Web of Science, 2015-2025) was performed. Examples of search terms include "artificial intelligence", "synthetic image generation", "autosegmentation", and "adaptive radiotherapy". Selected studies were peer-reviewed studies that describe, validate, or evaluate an AI tool for any online ART-related function in the domains of decision support, imaging, contouring, planning, and quality assurance (QA). KEY CONTENT AND FINDINGS: AI applications for online ART can be divided broadly into three domains: image registration, auto-contouring of tumor volumes and OARs, and ART decision support through dose and anatomical change prediction. New deep-learning based synthetic computed tomography (CT) generation using architectures like generative adversarial networks have enabled magnetic resonance imaging (MRI)-guided adaptive workflows. AI-driven deformable image registration (DIR) has also been shown to demonstrate alignment between planning and on-treatment imaging. Similarly, AI-driven enhancement of cone-beam CTs has improved their resolution, improving spatial information valuable for segmentation of tumor volumes. Auto-segmentation using U-Nets, ResNets, and one-shot learning has substantially reduced contouring time and approaches expert performance for many organs at risk (OARs), though challenges persist for small structures and evolving gross tumor volume (GTV)/clinical target volume (CTV) volumes. AI-based dose prediction and personalized plan optimization demonstrate clinically meaningful reductions in OAR dose, and early commercial systems such as Varian Ethos show promising real-world feasibility. Decision-support systems leverage these AI-workflows, using predictive modeling to guide daily dose adjustments. CONCLUSIONS: AI has the potential to transform ART, automating labour-intensive tasks while still achieving high-quality plans. To support safe clinical adoption, future work should emphasize multi-institutional validation, standardized benchmarking and contouring guidelines, and development of transparent, uncertainty-aware models.