2981 - Automatic Radiation Treatment Planning via Deep Learning-Based Dose Prediction for Left Breast Volumetric-Modulated Arc Therapy
Presenter(s)
J. Park1,2, D. Jung3,4, J. S. Chang3, S. H. Ahn5, J. S. Kim3, and J. Kim2; 1Department of Integrative Medicine, Yonsei University College of Medicine, Seoul, Republic of Korea, Seoul, Korea, Republic of (South), 2Department of Radiation Oncology, Gangnam Severance Hospital, Yonsei University College of Medicine, Seoul, Republic of Korea, Seoul, Korea, Republic of (South), 3Department of Radiation Oncology, Yonsei Cancer Center, Yonsei University College of Medicine, Seoul, Republic of Korea, Seoul, Korea, Republic of (South), 4Medical Physics and Biomedical Engineering Lab (MPBEL), Yonsei University College of Medicine, Seoul, Korea, Seoul, Korea, Republic of (South), 5Department of Radiation Oncology, Samsung Medical Center, Seoul, Republic of Korea, Seoul, Korea, Republic of (South)
Purpose/Objective(s):
This study aims to develop a deep learning-based dose prediction algorithm for left breast volumetric-modulated arc therapy (VMAT) and to investigate the feasibility of automatic radiation treatment planning (auto-planning).Materials/Methods:
A retrospective dataset of 100 left breast cancer patients was considered. We developed a deep learning network to predict a radiation dose distribution based on patient CT image and RT structures. This dose prediction algorithm was developed by combining a 2.5-D U-net network structure with a simple parameter-free attention module (SimAM). For the training and testing of the developed deep learning network, datasets were divided in ratios of 8:2 respectively. The radiation dose distribution predicted by the developed network was converted to a deliverable plan and corresponding radiation dose distribution using a dose mimicking algorithm in a commercial treatment planning system. The predicted and converted dose distributions were compared with those of the treatment plans (ground truth) which were delivered to the patients; we calculated differences in mean dose and V95% of planning target volume (PTV), mean heart dose, and mean lung dose. We further evaluated the accuracy of the predicted and converted dose distributions by calculating the Dice similarity coefficient (DSC) of the iso-volumes of 100% prescription dose.Results:
The prediction error (absolute difference between the ground truth and predicted values) of the developed network was 0.11 ± 0.15% for the mean PTV dose, 0.14 ± 0.50% for the PTV V95% dose, 0.30 ± 0.15% for the mean heart dose, and 0.19 ± 0.21% for the mean lung dose, respectively. The iso-volume DSC was 0.87 ± 0.05, and the deep learning network required less than 5 seconds to predict a three-dimensional dose distribution. When the predicted dose distribution was transformed into a deliverable treatment plan via the dose mimicking algorithm, the resulting dose distribution showed a 0.08 ± 0.12% difference compared to the original clinical plan.Conclusion:
In this study, the developed deep learning model based on 2.5-D U-net and SimAM successfully demonstrated high accuracy in the dose prediction for left breast VMAT. The rapid prediction time (<5 seconds) and the favorable dosimetric agreement between the clinical plans and auto-generated plans highlight the feasibility of clinically implementing dose prediction-based auto-planning workflow. Further research is warranted to refine the model and fully implement this approach into clinical practice, aiming to enhance both planning efficiency and consistency of radiotherapy.