SWFT-Net: a deep learning framework for efficient fine-tuning spot weights towards adaptive proton therapy

Area of Science:

• Medical Physics
• Radiation Oncology
• Artificial Intelligence in Medicine

Background:

• Adaptive proton therapy requires time-consuming spot weight re-tuning and plan reoptimization.
• Current methods are labor-intensive, hindering the transition to online adaptive proton therapy.

Purpose of the Study:

• To develop and evaluate a deep learning framework (SWFT-Net) for rapid adaptive proton therapy planning.
• To significantly reduce the time and human interaction needed for spot weight re-tuning and plan reoptimization.

Main Methods:

• A deep learning framework (SWFT-Net) was developed using residuals of dose maps and spot weights as inputs and outputs.
• Data augmentation involved modifying spot weights to create three datasets (DS10, DS30, DS50) for training and testing.
• Quantitative analyses included normalized root mean square error (NRMSE) of spot weights, Gamma passing rate, and dose difference within the planning target volume (PTV).

Main Results:

• SWFT-Net generated adapted plans in under a second on a GPU.
• NRMSE for spot weights ranged from 0.41% to 2.04% across datasets.
• Mean relative dose difference was consistently low (0.64%-0.92%), with Gamma passing rates over 95% for all datasets.

Conclusions:

• SWFT-Net demonstrates significant potential for realizing adaptive proton therapy by offering superior speed, accuracy, and robustness.
• The framework can serve as an alternative to existing spot fine-tuning algorithms, minimizing human interaction.
• This study provides a foundation for developing truly online adaptive proton therapy workflows incorporating daily anatomical changes.