ESTRO 2024 - Abstract Book

S4557

Physics - Machine learning models and clinical applications

ESTRO 2024

2729

Digital Poster

Machine learning model for patient-specific QA prediction in stereotactic radiosurgery

Simone A Buzzi 1,2 , Monica Bianchi 1,2 , Caterina Zaccone 1,2 , Andrea Bresolin 2 , Damiano Dei 2,3 , Pasqualina Gallo 2 , Francesco La Fauci 2 , Cristina Lenardi 1,4 , Francesca Lobefalo 2 , Lucia Paganini 2 , Sara Parabicoli 1,2 , Marco Pelizzoli 1,2 , Giacomo Reggiori 2,3 , Stefano Tomatis 2 , Marta Scorsetti 2,3 , Pietro Mancosu 2 , Nicola Lambri 2,3 1 Università degli Studi di Milano, Physics, Milan, Italy. 2 IRCCS Humanitas Research Hospital, Radiotherapy and Radiosurgery, Rozzano, Milan, Italy. 3 Humanitas University, Biomedical Sciences, Pieve Emanuele, Milano, Italy. 4 INFN, Physics, Milan, Italy

Purpose/Objective:

Stereotactic radiosurgery (SRS) is a non-invasive RT single fraction procedure that uses highly focused radiation beams to treat lesions in the brain. In particular, in case of multiple brain metastases using a single isocenter with multiple non-coplanar arc, highly conformal dose distributions with millimeter accuracy are required, at the cost of extreme modulation of a linac’s mechanical components. Therefore, patient-specific quality assurance (PSQA) is performed to ensure that SRS plans are delivered as intended. However, measurement-based PSQA represents a time consuming workload which, in case of failure, could potentially lead to delays in treatment plan delivery. In this study, a machine learning (ML) tree-based ensemble model was developed to predict the gamma passing rate (GPR) based on plan complexity metrics and plan parameters.

Material/Methods:

Four hundred-fifty-eight patients (i.e., 1604 arcs) from the internal database were selected. All patients were optimized using HyperArc (Varian) and underwent PSQA by Portal Dosimetry (Varian). The GPR analyses were performed automatically using 3%(global)/1 mm with 10% dose threshold, and 95% action limit criteria. Ten plan complexity metrics and 9 plan parameters were computed for each field and were used as input features for the ML models to correlate with the corresponding GPR of the field. A range of tree-based ensemble ML regression models and classifiers, including extra trees, random forest, AdaBoost , and XGBoost, was considered. The best combination of hyper-parameters was identified via cross-validation, and was then utilized to retrain the models on the entire training set. Due to the small number of fields which failed PSQA (24.9% with GPR<95%), a weight sampling strategy was implemented to mitigate performance imbalance. For the regression model, instances with lower GPR values were assigned higher weights based on the frequency of GPR values present in the training set. In the case of the classifier, a higher weight was assigned to the class of instances with GPR<95%. The performance of the extra trees was assessed on the test set, using MAE and absolute error statistics for the regression model, and specificity and sensitivity for the classifier.

Results: