End-to-end automated radiotherapy planning using transfer learning to overcome data scarcity

Why clinicians should know about it

• Picked for Medical Physics (paper of the day, 16 October 2026): End‑to‑end automated VMAT planning with transfer learning

Abstract

PURPOSE: To develop and benchmark 3D dose prediction algorithms for volumetric modulated arc therapy for head-and-neck (HN) and pancreatic cancer treatments, and to demonstrate the clinical feasibility of converting the predicted doses into deliverable plans. METHODS AND MATERIALS: Three deep learning models were pretrained on 3232 public plans and fine-tuned on institutional cohorts (256 HN and 72 pancreas patients). The accuracy of the fine-tuning strategy was compared with the baseline trained from scratch using only institutional data. In a prospective setting, the predicted dose was used as the planning goal for fully automated plan generation. Automated plans were benchmarked against clinical plans using Plan Quality Metrics. The deliverability of the fully automated plans was verified by Mobius ion-chamber patient-specific quality assurance (PSQA). RESULTS: Relative to training from scratch, transfer learning significantly accelerated convergence and minimized prediction error in scenarios with limited site-specific data. For HN cases, fine-tuning yielded a prediction error of 1.76 Gy. Notably, models fine-tuned on just 5% (n = 12) of the dataset achieved accuracy comparable to models trained from scratch using 50% (n = 116) of the training data. In the pancreas cohort, fine-tuning resulted in a mean absolute error of 1.35 Gy, representing a 14.6% reduction compared to the baseline trained solely on institutional data. When used for automated planning, these predicted 3D doses improved planning target volume homogeneity and organs at risk sparing (esophagus and trachea) in HN plans, while pancreas plans aimed to meet clinical goals with reduced dose to the spinal cord and right kidney. All automated plans were confirmed as clinically deliverable via PSQA. CONCLUSION: Compared with models trained from scratch on institutional data, transfer learning significantly reduced prediction error and improved robustness. The feasibility of converting predicted doses into clinical plans while maintaining plan quality through a fully automated process was demonstrated.

Abstract as published, via PubMed.