Multi-model study of fast VMAT segment dose calculation with deep learning

Area of Science:

• Medical Physics
• Radiotherapy
• Computational Imaging

Background:

• Deep learning (DL) enables photon dose calculation using Beam's Eye View (BEV) or patient coordinates.
• Evaluating DL model accuracy and speed across coordinate systems is crucial for clinical implementation.

Purpose of the Study:

• To compare the dose calculation accuracy and speed of various DL models under BEV and patient coordinate systems.
• To introduce lightweight DL models for efficient photon dose calculation.

Main Methods:

• Utilized CT scans and VMAT plans from 24 prostate cancer patients.
• Trained and tested five DL models (CNN-ConvLSTM, CNN-Mamba, DoTA, C3D, DeepDose-C3D) using Monte Carlo simulations for dose data.
• Assessed accuracy via gamma passing rates (γPR) and dose-volume histogram metrics; measured dose calculation times on GPUs.

Main Results:

• All models achieved high accuracy (≥91.0% for segment doses at 2%/3 mm, ≥99.0% for plan doses at 1%/3 mm).
• Lightweight BEV models (CNN-ConvLSTM, CNN-Mamba) demonstrated significantly faster calculation times (e.g., 5.5s and 6.2s per plan) compared to patient-coordinate models.
• BEV-based models showed more robust segment performance.

Conclusions:

• Both BEV and patient-coordinate DL methods provide accurate photon dose calculations.
• BEV approaches offer more robust segment dose prediction.
• CNN-ConvLSTM and CNN-Mamba are promising lightweight DL models for fast and accurate photon dose calculation in radiotherapy.