KEY POINTS
- This narrative review examines CBCT-based adaptive radiotherapy from image acquisition through segmentation, synthetic CT generation, dose reconstruction, replanning and automated adaptation decisions, with literature reviewed through July 2026.
- CBCT remains attractive because it is already integrated into many modern linacs, but routine adaptive use is constrained by poor soft-tissue contrast, scatter, noise, truncation artifacts and unstable Hounsfield units, all of which can affect contouring and dose calculation.
- Synthetic CT generation, deformable registration and AI-based image enhancement can improve dose-calculation suitability, but visually realistic images do not guarantee correct electron-density information. The review therefore emphasizes dosimetric rather than image-similarity validation.
- Clinical studies already demonstrate meaningful dosimetric gains in selected settings. Reported examples include approximately 38% reductions in bladder and rectum D50 during cervical adaptation, improved bladder-cancer target coverage, and a 23.3% improvement in pancreatic PTV prescription coverage with adaptation.
- Automated segmentation is increasingly feasible, but geometric metrics such as Dice or Hausdorff distance are insufficient by themselves. Contours near steep gradients may have clinically important dosimetric errors despite apparently good geometric agreement.
- Adaptation decision systems currently follow two broad strategies: rule-based triggers using predefined dosimetric or anatomical thresholds, and data-driven AI models predicting whether adaptation is required. A cervical-cancer model cited in the review achieved adaptation-decision accuracy of 0.855, compared with 0.795 for physician consensus.
- The authors argue that the realistic near-term model is a semi-automated closed loop: automated processing and risk assessment produce an “adapt” or “continue” recommendation with relevant dose metrics and quality flags, while a clinician retains the final decision.
- Fully automated adaptation requires more than segmentation and plan generation. Safe implementation needs integrated automated planning, independent dose verification, QA, failure detection, audit trails, multicentre validation and prospective evidence that dosimetric improvements translate into toxicity or tumor-control benefit.
CLINICAL TAKEAWAY
CBCT-based online adaptation is moving from technical feasibility toward workflow automation, but the unresolved question is increasingly when adaptation is actually necessary. Automated decision support may ultimately be more important than adapting every fraction, but human review remains essential until thresholds, failure modes and clinical benefit are validated prospectively.