Introducing matrix sparsity with kernel truncation into dose calculations for fluence optimization.

Area of Science:

• Medical Physics
• Computational Biology
• Artificial Intelligence

Background:

• Deep learning for radiation therapy planning demands extensive data and computational resources, leading to lengthy training times.
• Current commercial treatment planning systems create data pipeline bottlenecks for AI/ML algorithms requiring frequent dose calculations.
• Approximations in dose calculation can accelerate AI/ML model training without compromising treatment outcomes.

Purpose of the Study:

• To investigate the impact of matrix sparsity, achieved through kernel truncation, on radiation therapy dose calculations for AI/ML applications.
• To develop and evaluate a novel algorithm for fluence optimization that reduces computational complexity.

Main Methods:

• Implemented matrix sparsity by introducing kernel truncation to prune voxels from computationally intensive parts of the dose calculation.
• Developed a voxel discrimination approach to reduce the computational burden.
• Validated the dose calculation method against established systems (Eclipse) using water phantoms and patient anatomy, without heterogeneity corrections.

Main Results:

• Achieved gamma index passing rates of approximately 99% for uniform fluence beams and 98% for modulated fluence beams.
• Demonstrated a significant reduction in computation time and storage requirements, proportional to the square of the sparsity tolerance.
• Showcased a potential cost reduction exceeding 10x compared to dense calculations.

Conclusions:

• Matrix sparsity through kernel truncation offers a viable method for accelerating AI/ML-driven radiation therapy dose calculations.
• The developed approach significantly reduces computational resources, enabling faster and more complex treatment planning.
• This method allows for calculations previously infeasible on standard systems, paving the way for advanced AI/ML applications in radiotherapy.