Single shot full plan deep learning dose computation for radiation therapy using spherical harmonics.

Area of Science:

• Medical Physics
• Artificial Intelligence in Radiation Therapy
• Computational Dosimetry

Background:

• Accurate dose computation is crucial for effective radiotherapy planning.
• Increasing radiotherapy cases necessitate faster planning methods.
• Traditional physics-based dose calculation can be time-consuming and lack sufficient accuracy.

Purpose of the Study:

• To develop a novel, physics-informed deep learning AI method for rapid and accurate dose calculation.
• To address the speed and accuracy limitations of existing radiotherapy dose computation techniques.
• To enable high-accuracy dose calculation for complex clinical VMAT and IMRT plans.

Main Methods:

• A two-stage deep learning approach combining Beer-Lambert law with spherical harmonics.
• Utilizing an image-to-image neural network in the second stage for dose prediction.
• Extensive data generation and augmentation on 1641 clinical plans across three body sites, exceeding 100,000 training samples.

Main Results:

• The AI model achieved high accuracy, with average gamma pass rates of 99.1% (2%/2mm) and 94.4% (1%/1mm) across multiple body sites.
• Demonstrated exceptional speed, with a run-time of 1.6 seconds on an RTX 4090 GPU.
• Validated on diverse clinical VMAT and IMRT plans, showing reliable performance.

Conclusions:

• The proposed physics-informed deep learning method enables fast and highly accurate dose calculation for radiotherapy.
• This AI approach is suitable for both few-field and many-field treatment plans.
• Offers a significant advancement in radiotherapy planning efficiency and precision.