A nonvoxel-based dose convolution/superposition algorithm optimized for scalable GPU architectures.

Area of Science:

• Medical Physics
• Computational Biology
• Radiotherapy Technology

Background:

• Real-time adaptive planning and treatment in radiation therapy is computationally complex and often infeasible.
• Graphics processing units (GPUs) offer potential for accelerating computational performance and improving dose accuracy.
• Optimizing data structures and memory access is crucial for efficient GPU utilization in medical applications.

Purpose of the Study:

• To present a nonvoxel-based (NVB) approach for radiation therapy dose calculation on GPUs.
• To maximize computational and memory access efficiency and throughput on GPUs.
• To enable real-time adaptive planning and treatment through computational acceleration.

Main Methods:

• A ray-tracing mechanism restructures 3D CT data into a nonvoxel-based framework.
• Data is resampled during ray-tracing, making the algorithm independent of original CT resolution.
• An exhaustive parameter search optimized sampling parameters (zenithal, azimuthal, radial, ray spacing) for dose accuracy and speed.
• Gamma analysis (2%/2mm) assessed dose distribution accuracy using phantoms and clinical lung CT datasets.

Main Results:

• The NVB GPU algorithm demonstrated significant computational gains over CPU-based methods, independent of data size.
• Optimized parameters (8 zenithal/8 azimuthal angles, 1mm radial sampling, 2mm ray spacing) maintained >99% dose accuracy (gamma test).
• A total performance improvement factor of >175,000 was achieved compared to a CPU benchmark, and 20x compared to a voxel-based GPU method.

Conclusions:

• The nonvoxel-based convolution method offers substantial performance improvements over generic GPU implementations.
• Accuracy is maintained compared to CPU-computed dose distributions.
• This algorithm is a key advancement for developing tools for adaptive radiation therapy systems.