Toward Fully Automated Prostate HDR Brachytherapy Planning Using Deep Learning–Based Dose Prediction
Abstract
Purpose
This study aims to develop and evaluate a framework to generate fully automated prostate HDR brachytherapy plans using predicted dose from deep learning model, reducing planner dependence and planning time.
Methods
A deep learning–based dose prediction model was trained using 219 retrospective prostate HDR brachytherapy cases. A 3D-UNet architecture was implemented to predict patient-specific dose distributions using anatomical structure masks, dwell position information, and spatial distance maps as inputs. The predicted dose was converted into a clinically deliverable treatment plan by solving an linear inverse problem, in which dwell times were optimized using a non-negative least squares (NNLS) solver and a precomputed dose-rate matrix. To evaluate the dose-to-plan conversion, clinical dose distributions were converted back into deliverable plans and corresponding dosimetric metrics were compared. The fully automated planning pipeline was further evaluated on an independent test cohort of 28 patients by comparing dosimetric metrics between automated and clinical plans. In addition, prescription-dose isodose Dice similarity coefficients and voxel-wise mean absolute dose differences were calculated to assess spatial dose agreement.
Results
The predicted dose distributions showed strong agreement with clinical plans across target and organs at risk. The reverse optimization step successfully generated clinically deliverable dwell time distributions for all test patients. Dosimetric differences between automated and clinical plans were within clinically acceptable ranges, indicating that the predicted dose could be translated into deliverable HDR brachytherapy plans.
Conclusion
This study demonstrates the feasibility of a fully automated prostate HDR brachytherapy planning pipeline that integrates deep learning–based dose prediction with inverse optimization. The proposed framework generates clinically deliverable plans with minimal human intervention and provides a foundation for more efficient and streamlined HDR brachytherapy planning workflows.