Multi-Task Learning for Predicting VMAT Plan Deliverability and 3D Delivered Dose from Visualized Plan Parameters and 3D Plan Dose
Abstract
Purpose
Patient-specific quality assurance (PSQA) for volumetric modulated arc therapy (VMAT) is essential for verifying plan deliverability, while it remains resource-intensive and inefficient. We developed a multi-task deep learning framework that jointly predicts PSQA pass/fail result and the corresponding 3D dose distribution prior to PSQA tasks.
Methods
The patient cohort consisted of 162 left-breast two-arc VMAT plans prescribed to 26 Gy in 5 fractions, which were split into 142 and 20 cases for training and testing, respectively. The ratio of QA-passing to QA-failing plans was approximately 2:1 in the training set and 1:1 in the test set. Model inputs included two modalities: 1) DICOM RT plan-derived 3D image from multi-leaf collimator (MLC) positions and monitor unit (MU) value at each control point and 2) TPS-calculated 3D dose distribution. Model outputs were binary PSQA pass/fail label reflecting plan deliverability, and 3D delivered dose distribution reconstructed from machine log files. A dual-encoder 3D U-Net with a cross-attention module fused complementary input features for multi-task learning, using L1-loss for 3D dose prediction and cross-entropy for PSQA result prediction. The classification performance was evaluated using accuracy, F1-score, and area under the curve (AUC), and dose prediction performance was assessed by gamma passing rate (GPR) at the 3%/3 mm gamma criterion and structural similarity index measure (SSIM).
Results
The model yielded accuracy 0.80, F1-score 0.83, and AUC 0.93 for deliverability prediction, demonstrating strong discrimination between PSQA pass and fail plans. For dose prediction, it achieved GPR 98.45±2.96% and SSIM 0.98±0.01, indicating close agreement between the predicted and reference dose distributions. These outputs enable early identification of PSQA failure risk and facilitate downstream dose analysis via the predicted 3D dose distribution.
Conclusion
The proposed multi-task model integrates deliverability classification with 3D dose reconstruction, enabling identification of potentially failing plans prior to PSQA tasks.