Main Session
Sep
28
PQA 03 - Digital Health Innovation and Informatics, Patient Safety & Quality, and Radiation and Cancer Biology
2625 - Development of "DualGADE-Net" for Simultaneous Prediction of Gamma Passing Rates and Absolute Dose Errors In Patient-Specific QA
Presenter(s)
Hikaru Tanno, BS - Tohoku University, Sendai, Miyagi
H. Tanno1, N. Kadoya1, R. Tozuka1,2, S. Tomori3, T. Hoshino1, K. Arai1, Y. Katsuta1, and K. Jingu4; 1Department of Radiation Oncology, Tohoku University School of Medicine, Sendai, Japan, 2Department of Therapeutic Radiology, University of Yamanashi, Chuo, Japan, 3National Hospital Organization Sendai Medical Center, Sendai, Japan, 4Tohoku University Graduate School of Medicine, Sendai, Miyagi, Japan
Purpose/Objective(s): The purpose of this study was to develop "DualGADE-Net" (Dual Gamma and Absolute Dose Estimator Network) for the simultaneous prediction of gamma passing rates (GPR) and absolute dose errors (ADE) in patient-specific quality assurance for IMRT.
Materials/Methods:
108 prostate VMAT cases were analyzed. Inputs for the deep learning model, "DualGADE-Net," consisted of dose distributions calculated by TPS and MLC leaf position maps for each control point. The architecture of DualGADE-Net utilized a convolutional neural network backbone as a common base, featuring two independent output layers designed to compute GPR and ADE in parallel. To validate the effectiveness of the proposed model, STL models using the same network configuration were used. The mean absolute error (MAE) and Pearson’s correlation coefficient (r) were calculated.Results:
Regarding GPR, for STL, the MAE [%] at 2%/2mm and 3%/2mm were 2.98%±0.40% and 2.09%±0.28%, with r values of 0.56±0.10 and 0.54±0.12, respectively. For DualGADE-Net, they were 3.20%±0.26% and 2.19%±0.10% with r values of 0.46±0.11 and 0.47±0.17. Regarding ADE, the STL model exhibited significant overfitting with an extremely low correlation of 0.061±0.06. In contrast, DualGADE-Net achieved 0.26±0.18, confirming the mitigation of overfitting and the improvement in r.Conclusion:
We successfully developed "DualGADE-Net" for the simultaneous prediction of GPR and ADE, performing the first comparative validation against individual single-task models. Our results suggested that DualGADE-Net serves as a robust tool for enhancing the reliability of patient-specific QA and supporting high-standard dose verification in clinical practice.| Task | Metric | STL | DualGADE-Net | |
| gamma passing rates | MAE [%] | 2%/2mm | 2.98 ± 0.40 | 3.20 ± 0.26 |
| 3%/2mm | 2.09 ± 0.28 | 2.19 ± 0.10 | ||
| Pearson’s r | 2%/2mm | 0.56 ± 0.10 | 0.46 ± 0.11 | |
| 3%/2mm | 0.54 ± 0.12 | 0.47 ± 0.17 | ||
| absolute dose error | MAE [%] | 0.56 ± 0.11 | 0.46 ± 0.04 | |
| Pearson’s r | 0.061 ± 0.06 | 0.26 ± 0.18 | ||