A Convolutional neural Network Model for Predicting Gamma Passing Rate of 3D Array Detector-Based VMAT QA
T Matsuura1,2*, D Kawahara2, A Saito3, E Shiba4, K Yamada1, S Ozawa1,2, Y Nagata1,2, (1) Hiroshima High-Precision Radiotherapy Cancer Center, Hiroshima, Japan, (2) Hiroshima University, Hiroshima, Japan, (3) Hiroshima University Hospital, Hiroshima, Japan, (4) Department of Radiation Oncology, Hospital of the University of Occupational and Environmental Health, Fukuoka, Japan
Presentations
PO-GePV-T-188 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: The predicting models of gamma passing rate (GPR) via artificial intelligence have been proposed for improving the efficiency of the patient-specific quality assurance (QA) workloads. The current study aims to develop a convolution neural network (CNN) model for predicting GPR, and to propose a novel omitted workflow of prediction-based VMAT QA.
Methods: Two hundred and twenty-five VMAT plans measured GPR by ArcCHECK were used in this study. The dose and DUP distribution on the detector element plane were used as input data for the predicting model. The DUP distribution was generated by accumulated field edges weighted by a segmental monitor unit. Our CNN model comprised nineteen layers and was trained for 200 epochs. 10-fold cross-validation was applied to verify the predicting performance. The GPR values were predicted for thirty-five test cases, and the difference between the measured and predicted GPR values was evaluated. The omission rate of the measurement was calculated for cases with GPR 2SD of prediction error higher or lower from 90% set as the action limit at 3%/3-mm and 3%/2-mm tolerances.
Results: Mean absolute error (MAE) between measured GPR and predicted GPR in the CNN model using the dose distribution was 2.0% for 3%/3-mm, 3.1% for 3%/2-mm, and 6.6% for 2%/2-mm tolerances. By adding the DUP distribution to the CNN model, the MAE was improved to 1.8%, 3.0%, 6.0% for 3%/3-mm, 3%/2-mm, and 2%/2-mm tolerances, respectively. The effort of the measurements for patient‐specific QA could reduce by 83% and 23% for 3%/3-mm and 3%/2-mm tolerances, respectively.
Conclusion: The CNN model with the dose and DUP distributions on the detector enabled the prediction of GPR value with high accuracy. The proposed workflow with the CNN model could potentially help to omit the patient‐specific QA workloads.
ePosters
Keywords
Quality Assurance, Radiation Therapy, Treatment Verification
Taxonomy
TH- External Beam- Photons: Quality Assurance - IMRT/VMAT
Contact Email