Deep learning-based tools to distinguish plan-specific from generic deviations in EPID-based in vivo dosimetry

Area of Science:

• Medical Physics
• Radiation Oncology
• Artificial Intelligence in Healthcare

Background:

• Electronic portal imaging device (EPID)-based in vivo dosimetry (EIVD) is crucial for radiation therapy quality assurance.
• Deviations between planned and delivered doses, categorized as generic or plan-specific, complicate EIVD analysis.
• Traditional gamma-evaluation methods struggle to differentiate and manage these deviation types effectively.

Purpose of the Study:

• To evaluate the efficacy of deep learning (DL) tools in distinguishing generic deviations from plan-specific deviations in EIVD.
• To enhance the detectability of clinically relevant, patient-specific deviations during radiation therapy.

Main Methods:

• A 3D U-Net autoencoder was trained to identify patterns of generic deviations in gamma-distributions across various treatment sites (VMAT lung, prostate, head-and-neck; IMRT breast).
• The DL model's performance was compared against traditional gamma-analysis using receiver operator characteristic analysis for detecting introduced patient-related deviations (positioning, weight, tumor volume changes).
• Clinical relevance was assessed using 793 in vivo EIVD cases.

Main Results:

• The DL network demonstrated superior error detectability for patient-related deviations compared to gamma-analysis (average AUC 0.86 vs. 0.69).
• Significant differences in classification were observed between DL and gamma-analysis, particularly for head-and-neck (18%) and breast (64%) cancer treatments.
• DL successfully identified patient-related deviations missed by gamma-analysis and distinguished true alerts from those caused by EPID limitations or TPS inaccuracies.

Conclusions:

• DL-based tools can automate the identification of false positive gamma-alerts arising from generic deviations.
• These DL tools significantly improve the detection of critical plan-specific deviations, reducing the risk of false negatives.
• The proposed DL method offers substantial added value for managing large-scale EIVD systems in radiation therapy.