Image-based features in machine learning to identify delivery errors and predict error magnitude for patient-specific
Ying Huang1, Yifei Pi2, Kui Ma3
1Shanghai Chest Hospital, School of Medicine, Shanghai Jiao Tong University, 200030, Shanghai, China.
Summary
Machine learning accurately identifies radiation therapy delivery errors using image-based features. This approach predicts error types and magnitudes, enhancing quality assurance in intensity-modulated radiation therapy.
Area of Science:
- Medical Physics
- Radiotherapy Quality Assurance
- Machine Learning Applications
Background:
- Radiation therapy delivery errors can significantly impact treatment outcomes.
- Accurate identification and quantification of these errors are crucial for patient safety.
- Current quality assurance methods may not fully capture the nuances of delivery inaccuracies.
Purpose of the Study:
- To develop a machine learning model for identifying radiation therapy delivery error types.
- To predict the magnitude of associated delivery errors using image-based features.
- To enhance patient-specific quality assurance in intensity-modulated radiation therapy (IMRT).
Main Methods:
- Utilized 40 thoracic treatment plans with 208 beams, introducing four error types: collimator misalignment (COLL), monitor unit (MU) variation, systematic multileaf collimator misalignment (MLCS), and random MLC misalignment (MLCR).
- Extracted 14 image-based features from reference dose distributions (RDD) and error-introduced dose distributions (EDD) using portal dose image prediction (PDIP).
- Employed a random forest model for multiclass classification of error types and regression for error magnitude prediction.
Main Results:
- The top predictive features included relative displacement in the x-direction and various ratios of residual errors to the maximal RDD.
- Achieved high classification accuracy: 99.85% on validation and 99.30% on testing sets, with specific accuracies for error types ranging from 98.0% to 100%.
- Error magnitude prediction showed mean absolute errors (MAE) between 0.03-0.33 and root mean squared errors (RMSE) between 0.17-0.56 (validation), with similar ranges for the test set. Random MLC misalignment (MLCR) had the lowest prediction performance (70.1-96.6%).
Conclusions:
- Image-based features extracted from dose predictions are effective for identifying radiation therapy delivery errors.
- Machine learning models can accurately predict the magnitude of these errors.
- This approach integrates traditional gamma analysis with clinically relevant error classification and magnitude prediction for IMRT quality assurance.
