Deep learning-based detection and classification of multi-leaf collimator modeling errors in volumetric modulated...

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Artificial Intelligence in Medicine

Background:

• Volumetric Modulated Radiation Therapy (VMAT) relies on precise multi-leaf collimator (MLC) modeling parameters.
• Errors in MLC parameters like transmission factor (TF) and dosimetric leaf gap (DLG) can impact treatment accuracy.
• Accurate detection of these errors is crucial for patient safety and effective cancer treatment.

Purpose of the Study:

• To develop and assess deep learning (DL) models for detecting and classifying MLC modeling parameter errors (TF and DLG) in VMAT.
• To evaluate the performance of DL models in distinguishing between error-free plans and plans with specific TF or DLG errors.
• To investigate the models' ability to classify the sign (positive or negative) of TF and DLG errors.

Main Methods:

• Utilized 33 clinical VMAT plans for prostate and head-and-neck cancers in a phantom study.
• Introduced artificial errors (±10%, 20%, 30%) to TF and DLG parameters within the treatment planning system.
• Applied Gaussian filters to mimic measurement dose maps and generated dose difference maps for DL model training and evaluation.

Main Results:

• DL models successfully detected and classified TF and DLG errors with high accuracy for certain filter types ($$).
• Accuracy decreased significantly for other filter types ($$) and varied with error magnitude and treatment site.
• Models accurately classified the sign of errors for specific filter types ($$ and 1.0), while gamma analysis failed to detect errors.

Conclusions:

• Deep learning models offer a feasible approach for detecting and classifying TF and DLG errors in VMAT dose distributions.
• Model performance is influenced by error magnitude, treatment site, and the degree of simulated measurement dose.
• Further research can refine DL models for robust error detection in clinical VMAT applications.