Machine learning techniques for patient-specific quality assurance in Volumetric Modulated Arc Therapy (VMAT)
Loading...
Date
2024-10-25
Authors
Τσακιρίδου, Αγλαΐα
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Since the development of IMRT techniques in radiation therapy, comprehensive patient-specific quality assurance (PSQA) procedures have been conceived to verify the delivered dose. In volumetric modulated arc therapy (VMAT) the modulating parameters of a treatment plan are several, hence the delivery of the plans becomes more complex, and the confirmation of the accurate delivered dose is an important procedure. The main objective of this study is to assess the ability of machine learning (ML) and deep learning (DP) models to classify VMAT treatment plans as “pass” or “fail” based on plan complexity metrics, in purpose of streamlining the PSQA process and reducing the workload.
In this research, 99 patients with a total of 104 Head and Neck VMAT treatment plans were included. Eighteen plan complexity metrics were collected and calculated for each treatment plan and sixteen of them were used as input in the ML algorithms. Four classification models were employed, a Logistic Regression, a Random Forest, a Support Vector Machine (SVM) and a Convolutional Neural Network (CNN), to classify the results as “pass” or “fail”. The dataset was divided into an 80% training/validation set and a 20% test set in a random manner and a 10-fold cross validation was implemented to all four models. The classification performance was assessed by calculating the area under curve (AUC), the accuracy, the precision and the sensitivity of the models to recognize the “fail” plans. The receiver operating characteristic ROC curve for the first three classifiers, along with the accuracy and the loss for the performance of the CNN model, were plotted.
The AUC, the accuracy and the sensitivity for the “fail” class received the corresponding values for the model a) Logistic regression: 0.9417, 0.89 and 0.75, b) Random forest: 0.9944, 0.93 and 0.88, c) SVM: 0.9347, 0.91 and 0.79 d) CNN: 0.9361, 0.87 and 0.79. For the Logistic Regression and Random Forest models on account of their interpretability, a list of the three most important features in the classification process along with their weighting factor was created.
All four models seem to perform excellent classification task, since the AUC value is >0,9, which is also supported by the high values of sensitivity in “fail” class. Nevertheless, there is a concern about potential overfitting, which could be alleviated by selecting only the important features as ML input, utilizing a different dataset or/and collecting more data. The implementation of ML algorithms can enhance PSQA efficiency, by informing about the failure of a plan prior to measurements, thus saving time and resources.
Description
Keywords
Program Standards and Quality Assurance (PSQA), Volumetric Modulated Arc Therapy (VMAT), Complexity metrics, Classification, Logistic regression, Random forest, Support Vector Machine (SVM), Convolutional neural network (CNN)