Proton radiography interpretation with artificial intelligence for treatment deviation detection in proton therapy

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence

Background:

• Proton therapy dose conformity is challenged by patient setup errors, anatomical changes, and range uncertainties.
• Adaptive proton therapy (APT) aims to mitigate these issues through plan monitoring and adjustments.
• Proton radiography (PR) is a key technology for detecting range deviations in real-time.

Purpose of the Study:

• To develop and evaluate an artificial intelligence (AI) tool for automated interpretation and classification of treatment deviations using PR.
• To assess the tool's performance in detecting simulated and real-world deviations in proton therapy.

Main Methods:

• Synthetic CT data from 32 head-and-neck cancer patients were modified to simulate setup, calibration, and anatomical errors.
• Proton radiography simulations generated integral depth dose curves and range shift maps (RSMs).
• A convolutional neural network (EfficientNet-v2-M) was trained for multi-label classification of deviations on 14,503 RSMs and validated on clinical data from 22 patients.

Main Results:

• The AI tool achieved high performance on synthetic data: 97% precision, 92% recall, 93% F1-score, and 92% F2-score.
• On clinical data, the tool demonstrated strong generalization with 86% precision, 88% recall, 86% F1-score, and 87% F2-score.
• Classification of treatment deviations occurred within one second per image.

Conclusions:

• The AI-powered PR tool facilitates rapid, automated detection of treatment deviations in proton therapy.
• This technology supports integration into adaptive proton therapy (APT) workflows.
• The tool enhances quality assurance for online plan adaptation in proton therapy.