Deep learning techniques for proton dose prediction across multiple anatomical sites and variable beam...

Area of Science:

• Medical Physics
• Radiation Oncology
• Artificial Intelligence in Medicine

Background:

• Proton therapy offers precise radiation delivery, but accurate dose prediction is crucial for treatment planning.
• Artificial intelligence (AI) models are being developed to automate and improve dose prediction.
• Challenges remain in optimizing AI models for diverse patient data and complex treatment scenarios.

Purpose of the Study:

• To assess the impact of beam masks and data aggregation on AI-based dose prediction accuracy in proton therapy.
• To evaluate these strategies in scenarios with limited or heterogeneous patient datasets.
• To determine the optimal combination of beam masks and data aggregation for improved accuracy.

Main Methods:

• Trained convolutional neural networks on 541 prostate and 632 head and neck (H&N) proton therapy plans.
• Evaluated model performance using beam masks (radiation path depictions) and data aggregation techniques.
• Measured accuracy using dose-volume histogram (DVH) scores, mean absolute error, dice similarity coefficients (DSCs), and gamma passing rates.

Main Results:

• Beam mask inclusion improved dose prediction, particularly in low-dose regions and for varied beam configurations.
• Data aggregation alone yielded mixed results, improving high-dose accuracy but potentially degrading low-dose accuracy.
• Combining beam masks and data aggregation achieved the best overall performance, with significant improvements (up to 0.2 DSC) in heterogeneous H&N cases.

Conclusions:

• AI models integrating beam masks and data aggregation substantially enhance proton therapy dose prediction accuracy.
• This combined approach is particularly beneficial for complex cases and heterogeneous datasets.
• The findings suggest potential for accelerating proton therapy planning and improving cancer treatment efficacy.