Gamma Passing Rate Predictions Based On Automatic Feature Extraction of Modulation Maps and Monitor Unit Profiles: A Machine Learning Approach for Virtual Specific-Plan Verification
P Quintero1,2*, D Benoit1, Y Cheng1, C Moore2, A Beavis2, (1) University Of Hull, UK,(2) Queens Centre for Oncology & Haematology, Cottingham, UK
Presentations
PO-GePV-M-38 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: Various machine learning (ML) models have been proposed for gamma passing rates (GPRs) predictions to facilitate virtual specific-plan verification. These methods aim to avoid unnecessary measurements of plans predicted to get lower GPR (failing) values by immediately rejecting them. However, these models are mainly based on dose distribution data and manually extracted features such as modulation complexity metrics. We have implemented automatic feature-extraction models based on two linac parameters linked to each individual plan: the modulation map or leaf trajectories map (LTM) (2D array) and the delivered monitor units per control points profile (MU_cp) (1D array).
Methods: The area under the ROC curve (ROC-AUC) and prediction accuracy were calculated to evaluate the model performance of three convolutional neural network (CNN) models. A set of 1233 prostate VMAT plans was used. Model 1 was based on the MU_cp, model 2 on the LTMs, and (a hybrid) model 3 on both. Five-fold cross-validation was implemented to reduce overfitting, and the dataset splitting ratio for training, validation, and testing was 70%/20%/10%. A balanced training set of approximately equal failing plans (625) to passing plans (608) was used to train the models to perform binary classification of GPR at the 2%(global)/ 1mm criteria.
Results: The ROC-AUC values for model testing were 0.63 ± 0.08, 0.84 ± 0.06, and 0.89 ± 0.06 for model 1, 2, and 3 respectively; similarly, the model accuracies were 0.57 ± 0.08, 0.78 ± 0.05, and 0.81 ± 0.03 respectively. Additionally, the activation regions from LTM can be potentially included for modulation analysis and detect demanding MLC conditions.
Conclusion: Hybrid ML models based on plan features linked to treatment unit parameters are a feasible strategy to predict dose deliverability. This approach can generate more insight into features directly related to specific hardware configurations.
Keywords
Quality Assurance, DICOM-RT, Modeling
Taxonomy
TH- Dataset Analysis/Biomathematics: Machine learning techniques
Contact Email