Error detection and classification in patient-specific IMRT QA with dual neural networks

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Artificial Intelligence in Healthcare

Background:

• Gamma analysis is the standard for patient-specific quality assurance (QA) in intensity-modulated radiotherapy (IMRT).
• Gamma analysis has limitations in detecting small, clinically relevant errors and classifying error sources.
• Current methods lack efficiency in identifying the root cause of deviations in IMRT dose distributions.

Purpose of the Study:

• To propose a dual neural network (NN) method for simultaneous error detection and classification in patient-specific IMRT QA.
• To overcome the limitations of gamma analysis in sensitivity and error source identification.
• To develop a more active and informative approach to IMRT QA.

Main Methods:

• Extracted dose difference histogram (DDH) for low dose gradient regions and signed distance-to-agreement (sDTA) maps for high dose gradient regions.
• Utilized an artificial neural network (ANN) for DDH analysis and a convolutional neural network (CNN) for sDTA map analysis.
• Trained and validated networks on 13 IMRT plans (88 fields) with simulated errors, employing a fivefold cross-validation technique.

Main Results:

• The ANN achieved high accuracy in classifying monitor unit (MU) scaling and multileaf collimator (MLC) transmission errors (overall 98.3% ± 0.7%).
• The CNN demonstrated strong performance in classifying spatial errors, with specific accuracies generally above 90%.
• Distinct features in DDH and sDTA maps proved suitable for error classification, with overall CNN accuracy at 95.6% ± 1.5%.

Conclusions:

• The dual neural network method effectively detects and classifies errors in IMRT QA with excellent accuracy.
• DDH and sDTA maps are valuable features for robust error classification in IMRT QA.
• This approach can complement gamma analysis, shifting IMRT QA towards active error identification and root cause analysis.