CBCT-based adaptive radiotherapy and automated radiotherapy decision-making systems: development and current status.

Cone-beam computed tomography (CBCT)-based adaptive radiotherapy (ART) has emerged as a promising strategy for addressing inter-fractional anatomical and physiological changes during treatment. By incorporating on-treatment imaging, ART enables treatment plans to be adapted to the anatomy of the day, with the potential to improve target coverage and reduce dose exposure to organs at risk (OARs). However, routine implementation remains constrained by the limited image quality of CBCT, uncertainties in dose calculation, labour-intensive workflows and the lack of reliable criteria for determining when adaptation is clinically necessary. This review summarises the development, current status and future prospects of CBCT-based ART, with particular emphasis on artificial intelligence (AI)-assisted automated delineation and automated radiotherapy decision-making systems. CBCT is widely available on modern linear accelerators and remains the most practical imaging basis for ART in many centres. Nevertheless, poor soft-tissue contrast, noise, artefacts, field-of-view truncation and unstable Hounsfield unit values limit its performance in target delineation and dose reconstruction. Recent advances in image-quality enhancement, deformable image registration and synthetic CT generation have partially addressed these limitations. AI-based auto-segmentation models, including convolutional neural networks, Transformers, generative adversarial networks and diffusion models, have demonstrated increasingly favourable geometric accuracy and workflow efficiency. Automated decision-making systems can advise clinicians whether treatment-plan adaptation is required and how it should be implemented according to each patient's circumstances, representing a prerequisite for more efficient ART. Current research has mainly focused on two approaches: adaptation triggered by predefined dosimetric thresholds, and artificial-intelligence-based models that learn from prior knowledge and predict potential dosimetric changes for the current patient based on previous experience. Both approaches have their respective strengths and limitations, but require further refinement, including the systematic exploration of decision thresholds. Moreover, the realisation of fully automated adaptive radiotherapy will require the integration of additional systems, including automated treatment planning and automated QA. Overall, CBCT-based ART assisted by AI is technically feasible and may improve dosimetric quality and treatment efficiency. Future progress requires improved CBCT-specific algorithms, integrated automated planning and QA, multicentre validation, standardised safety frameworks and stronger evidence of clinical benefit.