Parallel beamlet dose calculation via beamlet contexts in a distributed multi-GPU framework.

Area of Science:

• Medical Physics
• Computational Biology
• Radiotherapy

Background:

• Dose calculation is critical but computationally intensive in radiation therapy planning.
• Current methods struggle with the demands of automatic beam orientation and arc trajectory optimization.
• Existing graphical processing unit (GPU) methods parallelize single beamlets, under-utilizing GPU potential.

Purpose of the Study:

• To develop an efficient framework for parallel computation of numerous beamlet doses.
• To enable faster and more scalable dose calculations for advanced treatment planning.
• To address the need for accommodating hundreds to thousands of beam candidates simultaneously.

Main Methods:

• A novel context-based transformation separates beamlet density and TERMA for independent dose calculation.
• Beamlet contexts are processed in a composite array using GPU collapsed-cone convolution superposition.
• A distributed manager-worker architecture scales the approach across multiple GPUs.

Main Results:

• The context-based method achieves <1.35% average error compared to existing algorithms and Monte Carlo simulations.
• Demonstrates speedups of up to two orders of magnitude over sequential GPU methods.
• Exhibits near-linear scaling with the number of compute nodes and GPUs, ensuring flexibility.

Conclusions:

• The context-based approach sets a new performance standard for beamlet-based dose calculation.
• Accelerates large-scale treatment planning tasks, including 4π IMRT and VMAT, within clinical timeframes.
• Offers a flexible and powerful solution for various large-scale treatment planning challenges.