Development and Multi-Institutional External Validation of an Automatic IMRT Planning System for Highly Accurate Beam Delivery Using Deep Learning-Predicted Fluence Maps
Abstract
Purpose
High beam delivery accuracy is essential for reliable IMRT delivery and efficient QA. This study aimed to develop a deep learning-based automatic planning system to improve delivery accuracy using predicted fluence maps and to evaluate its performance and generalizability through internal and external validation without model retraining.
Methods
A deep learning-based fluence prediction model implemented using a generative adversarial network was developed using 101 prostate IMRT plans from a single institution, with the top 50% of plans based on gamma passing rate (GPR) selected for training. The predicted fluence maps were imported into a treatment planning system, where leaf motion and dose distributions were calculated to generate deliverable synthetic plans (SPs). Internal validation was performed using 20 independent cases by comparing SPs with corresponding clinical plans (CPs) in terms of EPID-based GPR and dose–volume metrics. For external validation, the trained model was applied to datasets from two external institutions (20 cases; 10 per institution) without model retraining, and SPs were evaluated using the same criteria.
Results
Both internal and external validations showed significantly higher GPRs for SPs than CPs. In the internal validation, GPRs increased by 1.8%, 3.8%, 6.4%, and 11.6% for the 3%/2-mm, 2%/2-mm, 2%/1-mm, and 1%/1-mm tolerances, respectively. In the external validation, GPRs increased by 1.5%, 2.6%, 7.3%, and 7.5%, respectively. In dosimetric evaluation, target coverage was slightly reduced in SPs, while most cases satisfied the dose constraints. All internal cases and 90% of external cases satisfied the target and organ-at-risk dose criteria.
Conclusion
The automatic IMRT planning system designed to improve beam delivery accuracy was developed and evaluated using internal and external datasets. The system improved beam delivery accuracy without model retraining and achieved clinically acceptable dosimetric performance in most cases. This approach may contribute to more efficient QA workflows and improved radiotherapy delivery accuracy.