Deep learning-based voxel sampling for particle therapy treatment planning.

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Scanned particle therapy planning is computationally intensive, particularly for heavy ions due to variable relative biological effectiveness.
• Current voxel selection strategies for treatment optimization often require manual input and lack efficiency.

Purpose of the Study:

• To develop and validate a novel deep-learning model for optimizing voxel selection in scanned particle therapy planning.
• To improve computational efficiency and reduce treatment planning time without compromising plan quality.

Main Methods:

• A deep-learning model based on P-Net was designed for automatic, optimal voxel sampling.
• The model was trained and tested on 70 head and neck patient plans receiving carbon ion therapy, utilizing an AI infrastructure.
• A custom loss function was implemented to balance sampling density and voxel reduction.

Main Results:

• The AI model reduced the number of voxels by a median of 84.8% with less than 1% loss in plan quality.
• Optimization time decreased by a factor of 7.5 for the full AI selection model.
• A targeted model reducing voxels in the target only achieved 71.6% reduction with 0.5% quality loss.

Conclusions:

• The proposed deep-learning voxel sampling technique significantly reduces computational time in scanned particle therapy.
• This method offers a negligible loss in plan quality, making it suitable for complex treatments and future real-time adaptation strategies.