Predicting the error magnitude in patient-specific QA during radiotherapy based on ResNet

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Radiotherapy quality assurance (QA) relies on accurate patient-specific dosimetry to evaluate treatment plan delivery.
• Error magnitude prediction is crucial for assessing radiotherapy plan delivery accuracy.
• No prior studies have explored deep learning for predicting radiotherapy delivery error magnitudes.

Purpose of the Study:

• To develop and evaluate deep learning models, specifically ResNet, for predicting the magnitude of various radiotherapy delivery errors.
• To assess the feasibility of using deep learning to quantify errors in intensity-modulated radiation therapy (IMRT).

Main Methods:

• Utilized 34 chest cancer IMRT plans (172 fields), with 30 plans for training/validation and 4 for testing.
• Introduced specific delivery errors: collimator misalignment (COLL), monitor unit variation (MU), random multi-leaf collimator shift (MLCR), and systematic MLC shift (MLCS).
• Employed ResNet with different input combinations (dose difference, gamma distribution, reference dose distribution + error-introduced dose distribution) to predict error magnitudes.

Main Results:

• Achieved high prediction accuracy for error types: 98.36% (dose difference), 98.91% (gamma distribution), and 100% (RDD+EDD).
• Demonstrated low Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) across all error types, particularly with RDD+EDD input (RMSE: 0.15-0.24, MAE: 0.11-0.18).
• Established ResNet-based models for predicting error magnitudes for COLL, MU, MLCR, and MLCS.

Conclusions:

• Successfully developed ResNet-based models for predicting radiotherapy delivery error magnitudes.
• The models provide accurate predictions for different error types, offering valuable insights for patient-specific QA.
• This deep learning approach can enhance the evaluation of radiotherapy plan delivery accuracy.