Prediction of VMAT gamma passing rates using 3D CNNs based on leaf position analysis and gradient class activation...

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Artificial Intelligence in Healthcare

Background:

• Volumetric Modulated Arc Therapy (VMAT) requires rigorous quality assurance (QA) due to its technical complexity.
• Current Gamma Passing Rate (GPR) measurements are resource-intensive.
• Existing predictive metrics for VMAT QA often lack accuracy due to beam model independence.

Purpose of the Study:

• To predict the Gamma Passing Rate (GPR) of VMAT arcs using a 3D convolutional neural network (3D CNN).
• To enhance the interpretability of VMAT QA predictions using Gradient-weighted Class Activation Mapping (Grad-CAM).
• To identify potential limitations in beam models and Linear Accelerator (LinAc) performance.

Main Methods:

• Trained a 3D CNN on 140 6 MV VMAT arcs from an Elekta Harmony Pro LinAc.
• Utilized discretized beam's eye view (BEV) representations and monitor unit (MU) values as input data.
• Evaluated GPR using a Delta4+ phantom (1%/2 mm criterion) and applied Grad-CAM for visualization.

Main Results:

• The 3D CNN model achieved a mean absolute error (MAE) of 2.0% on the test set.
• Grad-CAM analysis revealed that dynamic delivery aspects significantly influenced prediction accuracy.
• Model performance was robust with both standard and cropped input data.

Conclusions:

• Deep learning enables automated and efficient GPR prediction for VMAT QA.
• The model provides a scalable solution for ensuring treatment safety, particularly in time-critical adaptive radiotherapy.
• Grad-CAM offers valuable insights into beam model and LinAc performance, aiding in treatment planning and QA refinement.