人工知能による陽子線画像解釈を用いた陽子線治療における治療逸脱検出
Giuliano Perotti Bernardini1, Arthur Galapon1, Gabriel Guterres Marmitt1
1Department of Radiation Oncology, University Medical Center Groningen, University of Groningen, the Netherlands.
Background And Purpose:
Patient setup errors, anatomical changes, and uncertainties in proton range estimation degrade dose conformity in proton therapy. Adaptive proton therapy (APT) mitigates these deviations by monitoring and adjusting treatment plans. Proton radiography (PR) offers direct proton range information, making it a promising method to detect such deviations. In this study, we developed and evaluated an artificial intelligence (AI) PR tool for automated interpretation and classification of treatment deviations.
Materials And Methods:
Computed Tomography (CT) scans from 32 head-and-neck cancer patients were synthetically modified to simulate setup errors (±2-4 mm), calibration curve errors (±3-5% for fat/soft tissue, ±7-11% for bone), and anatomical changes (±2-12 mm mimicking weight variations). PR simulations were performed using OpenREGGUI to generate integral depth dose (IDD) curves and range shift maps (RSMs) across 260 × 260 mm2 PR fields, resulting in 14,503 RSMs. A convolutional neural network (EfficientNet-v2-M) was trained from scratch for multi-label classification. Performance was evaluated on the synthetic dataset and an independent clinical dataset of 22 patients who underwent plan adaptation.
Results:
The CNN classified treatment deviations within one second per image. On the synthetic dataset, it achieved 97% precision, 92% recall, 93% F1-score, and 92% F2-score. On the clinical validation set, it maintained high performance metrics: 86% precision, 88% recall, 86% F1-score, and 87% F2-score, demonstrating strong generalization to clinical scenarios.
Conclusions:
The AI-enhanced PR tool enables fast, automated detection of treatment deviations in proton therapy, supporting its integration into APT workflows and online plan adaptation for improved quality assurance.
