Towards An Ultra-Fast Deep-Learning-Based Dose Engine for Prostate VMAT Using Knowledge Distillation
W Tseng1*, H Liu2, Y Yang2, C Liu1, B Lu1, (1) Department of Radiation Oncology, University of Florida, Gainesville, FL, (2) Department of Industrial and Systems Engineering, University of Florida, Gainesville, FL
Presentations
PO-GePV-T-297 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: A relatively lower accurate (but fast) dose engine is typically chosen for a practical planning time, which could yield a sub-optimal plan and affect the success of radiotherapy. Therefore, to maintain high plan quality while shortening the overall planning time, we proposed a deep-learning-based dose engine for prostate VMAT that aims to achieve ultra-fast calculation speed with high accuracy via knowledge distillation (KD).
Methods: Our KD-based training framework consists of two models: a large pre-trained teacher (3D U-Net) and a lightweight to-be-trained student (variant 3D U-Net). During the training phase, student receives knowledge transferred from the pre-trained teacher for a better generalization. The trained student as a final dose engine can perform dose calculation independently. For each VMAT arc, the identical input to both models is low-accuracy estimated dose on water and patient CT images, and the output is predicted patient dose. The ground truth was patient dose computed by a treatment planning system (TPS, Pinnacle³). Twenty and ten prostate cancer patients were included for model training and assessment, respectively. The model’s performance (teacher, student with/without KD) was evaluated using 2%/2mm Gamma Passing Rate (GPR).
Results: The GPR between the predicted and TPS doses was 98.13±1.07% for student with KD and 98.89±0.79% for teacher, whereas it was reduced to 95.23±3.26% for student without KD. The inference time per arc was 0.013s (student) and 0.023s (teacher) on average.
Conclusion: The proposed dose engine can accurately and rapidly calculate patient dose for prostate VMAT. It offers a much faster calculation speed than traditional dose engines, and therefore it is a promising tool for accelerating the overall planning in online adaptive workflow where fast dose calculation is much desired. Since its lightweight size, our dose engine can be easily implemented on a limited computing device for a prompt dose calculation.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email