Toward automatic beam angle selection for pencil-beam scanning proton liver treatments: A deep learning-based...

Area of Science:

• Medical Physics
• Radiation Oncology
• Artificial Intelligence

Background:

• Proton therapy offers dose deposition advantages over photon therapy.
• Proton treatment planning is complex, with beam angle selection being critical and time-consuming.
• Automated methods for beam angle selection are needed to improve efficiency and consistency.

Purpose of the Study:

• To develop and evaluate a deep learning-based approach for automatic beam angle selection in proton pencil-beam scanning for liver lesions.
• To address challenges in beam angle selection, such as angular discontinuity.

Main Methods:

• A deep learning model was developed, treating beam angle selection as a multi-label classification problem.
• Novel techniques, including Circular Earth Mover's Distance regularization and circular-smooth labels, were used to handle angular discontinuity.
• An analytical algorithm was employed for post-processing to refine the AI-selected angles, mimicking clinical practice.

Main Results:

• AI-selected beam angles closely matched human planners' selections, with a median angular difference of 10°.
• 68% of AI-predicted beam angles were within 10° of human-selected angles.
• Proton treatment plans generated with AI-selected angles demonstrated comparable dosimetric quality to those using human-selected angles.

Conclusions:

• The study demonstrates the feasibility of a novel deep learning-based beam angle selection technique for proton therapy.
• AI-selected beam angles resulted in clinically viable treatment plans with comparable dosimetric outcomes for liver cancer patients.
• This AI model, combined with other automated tools, could pave the way for nearly fully automated proton therapy treatment planning.