Related Experiment Video
Updated: May 7, 2026
Author Spotlight: UAV Remote Sensing for Efficient Invasive Plant Biomass Estimation
Published on: February 9, 2024
Prediction of VMAT gamma passing rates using 3D CNNs based on leaf position analysis and gradient class activation
Johannes Berchtold1, Sara Vockner1, Ivan Messner1
1Department of Radiation Therapy and Radiation Oncology, Paracelsus Medical University, Salzburg, Austria.
Background:
Volumetric Modulated Arc Therapy (VMAT) is a highly conformal radiotherapy technique that enables precise tumor irradiation while sparing surrounding healthy tissue. However, the high technical demands this technique places on Linear Accelerators (LinAc) necessitate reliable quality assurance (QA) tools. The Gamma Passing Rate (GPR), commonly used to compare planned and delivered dose distributions, requires extensive measurement resources. Many existing predictive metrics, such as the popular Modulation Complexity Score (MCS), are independent of the beam model, limiting their accuracy. Consequently, identifying appropriate metrics and their individual thresholds can be challenging.
Purpose:
This study aims to predict the GPR of VMAT arcs using a three-dimensional convolutional neural network (3D CNN). Gradient-weighted Class Activation Mapping (Grad-CAM) is applied to improve interpretability, identify weak segments, and potentially reveal beam model limitations.
Methods:
A 3D CNN was trained on 140 6 MV VMAT arcs, with 30 arcs each used for validation and testing. All plans were delivered by an Elekta Harmony Pro LinAc with 4° control point (CP) spacing. Input data included discretized beam's eye view (BEV) representations and segment-specific monitor unit (MU) values. GPR evaluation was performed using a Delta4+ phantom with a 1%/ 2 mm criterion. Data augmentation enhanced training diversity. Grad-CAM was used to visualize influential plan regions.
Results:
After 36 epochs, the model achieved a mean absolute error (MAE) of 2.0%(test set) and 1.3%(training set). With cropped input, the best MAEs were 2.1%(test) and 1.5%(training). Grad-CAM analysis indicated that dynamic delivery aspects had more influence on prediction accuracy than static features like field shape.
Conclusions:
This study highlights the potential of deep learning for automated GPR prediction, offering a more efficient QA workflow. Especially in time-critical settings like online adaptive radiotherapy, where traditional measurement-based QA is often impractical, this model provides a scalable solution to ensure treatment safety. The use of Grad-CAM enables insight into beam model and LinAc performance, allowing refinement of treatment planning and improved QA precision in clinical practice.
