Detecting MLC modeling errors using radiomics-based machine learning in patient-specific QA with an EPID for...

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Machine Learning in Healthcare

Background:

• Intensity-modulated radiation therapy (IMRT) requires precise multileaf collimator (MLC) function for accurate dose delivery.
• MLC modeling errors, such as transmission factor (TF) and dosimetric leaf gap (DLG) variations, can compromise treatment efficacy and patient safety.
• Patient-specific quality assurance (QA) is crucial for verifying IMRT plan accuracy.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for detecting MLC modeling errors using radiomic features from electric portal imaging device (EPID)-measured fluence maps.
• To compare the performance of ML models against conventional gamma analysis for identifying specific MLC errors.

Main Methods:

• Fluence maps from 38 IMRT beams were analyzed, with simulated MLC errors (TF, DLG, positional) introduced.
• Radiomic features (837) were extracted from fluence difference maps.
• ML models (decision tree, kNN, SVM, logistic regression, random forest) were trained for binary classification of error-free vs. error plans.
• Fourfold cross-validation and a dedicated test set were used for model evaluation.

Main Results:

• ML models achieved high sensitivity and specificity for detecting individual MLC errors (TF, DLG, positional).
• Highest sensitivities: 0.913 (TF), 0.978 (DLG), 1.000 (positional).
• Highest specificities: 1.000 (TF), 1.000 (DLG), 0.909 (positional).
• Gamma analysis demonstrated lower sensitivity (0.737-0.882) compared to ML models.

Conclusions:

• Radiomics-based ML models offer superior sensitivity and specificity for detecting single types of MLC modeling and positional errors compared to gamma analysis.
• While effective for single errors, further model refinement is needed for detecting multiple concurrent errors.
• Radiomics-based IMRT QA presents a promising advancement for identifying MLC modeling inaccuracies.