Integrating aperture shape controller and machine learning prediction to improve gamma passing rates in lattice radiotherapy.
Authors
Affiliations (2)
Affiliations (2)
- Department of Radiotherapy, Shuguang Hospital Affiliated to Shanghai University of Chinese Traditional Medicine, Radiation Oncology Department No. 528, Zhangheng Road, Pudong New Area, Shanghai, Please select, 201203, CHINA.
- Department of Radiotherapy, Shanghai Shuguang Hospital Affiliated to Shanghai University of Traditional Chinese Medicine, Radiation Oncology Department No. 528, Zhangheng Road, Pudong New Area, Shanghai, Please select, 200021, CHINA.
Abstract
This study proposes a workflow integrating the Aperture Shape Controller (ASC) in the Varian Eclipse system with a machine learning-based verification prediction model to improve gamma passing rates (GPRs) of LATTICE Radiotherapy (LRT) plans and reduce repeated verifications.
Approach: LRT plans were generated on CT images from 20 patients (five each with head and neck, thoracic, abdominal, and bone metastases) using a Varian Halcyon v3.0. Twenty IMRT-based Non_ASC_LRT plans (nine fields each, 180 fields total) were created and re-optimized with ASC set to "Very High," producing ASC_LRT plans. All plans underwent portal dosimetry. Plan complexity features were extracted from DICOM files, and dosimetric parameters from DVHs. An eXtreme Gradient Boosting (XGBoost) classifier was trained 5 times with cross-validation on a balanced dataset of 196 fields generated by under-sampling and SMOTE, and subsequently applied to failed fields in Non_ASC_LRT plans to evaluate its potential guidance for ASC application.
Main results: ASC optimization improved GPR in fields with baseline values <98% (mean increase 1.95%, p<0.05), while overall GPR showed no significant change. Dosimetric parameters (Dmax, Dmean_target, Dmean_vertex, and valley-to-peak dose ratio (VPDR)) were not significantly affected, though VPDR showed a slight overall increase. Over five training rounds, the XGBoost model achieved an average accuracy of 94.38%, sensitivity of 94.60%, specificity of 94.13%, precision of 93.57%, and an F1-score of 93.98%. All six failed fields in Non_ASC_LRT plans were correctly identified, and five passed verification after ASC optimization, reducing re-planning.
Significance: Combining ASC with a prediction model effectively balances dosimetric quality and deliverability in LRT, enhances GPR in low-performing fields, and reduces verification workload, thereby improving clinical efficiency.