Prescription‑dose stratification improves deep learning‑based VMAT dose prediction in locally advanced NSCLC

Area of Science:

• Radiation Oncology
• Medical Physics
• Artificial Intelligence in Medicine

Background:

• Volumetric-modulated arc therapy (VMAT) planning for non-small cell lung cancer (NSCLC) is complex, requiring iterative optimization to balance tumor coverage and organ-at-risk (OAR) sparing.
• Deep learning (DL) offers potential to accelerate VMAT planning through dose prediction, but the effect of mixed radiation dose prescriptions during DL model training is not well understood.

Purpose of the Study:

• To evaluate whether stratifying DL models by radiation dose prescription improves VMAT dose prediction performance for NSCLC.
• To compare the accuracy of single-prescription DL models versus a mixed-prescription DL model for VMAT planning.

Main Methods:

• Seventy-two NSCLC VMAT cases were recalculated to 50, 54, and 60 Gy.
• Four 3D U-Net DL models were trained: three single-prescription models (50 Gy, 54 Gy, 60 Gy) and one mixed-prescription model (50 Gy + 60 Gy).
• Model performance was assessed using mean absolute error (MAE) for planning target volume (PTV) and OAR dose metrics.

Main Results:

• Single-prescription models achieved low MAEs for PTV coverage (<4 Gy) and hot-spot errors (<1 Gy), with mean OAR dose errors ≤2.3 Gy.
• The mixed-prescription model exhibited significantly larger errors, including PTV hot-spot MAE of 11.3 Gy and spinal cord maximum dose errors of 5-6 Gy.
• Voxel-wise analysis revealed localized dose deviations in low-dose lung regions and near steep dose gradients for the mixed model.

Conclusions:

• Prescription-dose stratification demonstrably improves clinically relevant VMAT dose prediction metrics for NSCLC.
• Findings support the use of prescription-stratified DL models as a valuable tool for radiation therapy planning decision support and optimization guidance.