Performance assessment of variant UNet-based deep-learning dose engines for MR-Linac-based prostate IMRT plans

Area of Science:

• Medical Physics
• Radiotherapy
• Deep Learning

Background:

• UNet-based deep learning (DL) architectures show promise as dose engines for linear accelerator (Linac) models.
• Existing studies use diverse strategies, hindering fair comparison of UNet-based engine performance.
• Magnetic-resonance (MR)-Linac systems require accurate dose calculations for intensity-modulated radiation therapy (IMRT).

Purpose of the Study:

• To comprehensively evaluate the performance of various UNet-based DL models for MR-Linac IMRT dose calculations.
• To establish a standardized comparison of different UNet architectures using consistent input data and environments.
• To assess both accuracy and computational efficiency of DL models in MR-Linac treatment planning.

Main Methods:

• Implemented and tested six UNet-based models: standard-UNet, cascaded-UNet, dense-dilated-UNet, residual-UNet (Res-UNet), HD-UNet, and attention-aware-UNet.
• Used patient CT and IMRT field dose in water as input, generating patient dose output.
• Compared DL-calculated doses against Monaco Monte Carlo (MMC) calculations as the reference standard, using gamma analysis for accuracy and inference time for efficiency.

Main Results:

• All evaluated UNet models accurately converted low-accuracy doses in water to high-accuracy patient doses within the MR-Linac's magnetic field.
• Gamma analysis showed high passing rates: >86% (1%/1 mm), >98% (2%/2 mm), and >99% (3%/3 mm) for all models.
• Inference times varied significantly (0.03s on GPU to 7.5s on CPU), with Res-UNet showing highest accuracy and standard-UNet the fastest inference.

Conclusions:

• UNet-based DL models are feasible for MR-Linac IMRT dose calculations.
• Model selection depends on balancing accuracy requirements with available computational resources.
• This study provides a fair assessment of UNet model performance under consistent conditions for MR-Linac applications.