A non-voxel-based broad-beam (NVBB) framework for IMRT treatment planning

Area of Science:

• Medical Physics
• Radiation Oncology
• Computational Imaging

Background:

• Current intensity-modulated radiation therapy (IMRT) optimization relies on voxel-based beamlet superposition (VBS), demanding substantial computational resources and storage.
• The VBS framework's temporal and spatial complexity limits large-scale IMRT planning efficiency and increases costs.

Purpose of the Study:

• To introduce a novel non-voxel-based broad-beam (NVBB) framework for IMRT planning.
• To enable large-scale IMRT planning with reduced computational resources, improved cost-effectiveness, plan quality, and planning throughput.
• To demonstrate the framework's efficient implementation on a graphics processing unit (GPU).

Main Methods:

• Developed a non-voxel-based broad-beam (NVBB) framework enabling direct treatment parameter optimization (DTPO).
• Evaluated objective functions and derivatives using a continuous viewpoint, eliminating the need for voxel and beamlet representations.
• Implemented the NVBB framework on a single workstation with a GPU (NVBB-GPU) and integrated it with the TomoTherapy treatment planning system (TPS).

Main Results:

• The NVBB framework exhibits linear complexities (O(N(3))) in space and time, with low memory requirements.
• Comparative benchmarks against a commercial VBS-based TPS running on a 14-node cluster showed comparable dose accuracy (within 1%) and plan quality.
• The NVBB-GPU demonstrated a significant reduction in planning time (by several folds) compared to the VBS-cluster, with superior target uniformity in some cases.

Conclusions:

• The novel NVBB framework facilitates efficient, large-scale IMRT optimization with improved plan quality and reduced costs.
• DTPO and the continuous viewpoint eliminate beamlet pre-calculation, leading to enhanced efficiency and potential for better treatment plans.
• GPU parallelization of the NVBB framework on a single workstation offers a cost-effective and faster alternative to traditional cluster-based VBS methods.