Examination of New Indicators in Adaptive Radiotherapy Using Machine Learning in Head and Neck Volumetric Modulated Arc Therapy

Abstract

<h3>Purpose/Objective(s)</h3> During head and neck intensity-modulated radiotherapy treatment, a decrease in the patient's bodyweight and tumor size can often lead to inconsistencies in fixation and over- or under-dosage around the target. Treatment accuracy can be ensured by image-guided radiotherapy (IGRT). If substantial changes occur in the tumor volume or bodyweight, adaptive radiotherapy (ART) is performed. There is no clear guideline to initiate ART, and the decision is largely operator-driven. This study aimed to propose a new ART index using cone-beam computed tomography (CBCT) for existing patient data using machine learning. <h3>Materials/Methods</h3> The clinical target volume (CTV) and the parotid glands on the planning CT were automatically contoured to CBCT by deformable image registration (DIR), and the dice similarity coefficient (DSC) according to the number of fractions that were calculated from the original ROIs and DIR_ROIs. These DSCs were normalized by the maximum value and differentiated between patients who underwent ART (n = 416) and those who did not (n = 704). To determine the ART trigger from these DSCs, we devised a program that could guide the decision boundary by machine learning. The decision boundary was calculated using Python 3.7.6 with a support vector machine (SVM). T-test was performed to observe whether there was a difference in the slopes of the CTV and the parotid glands with and without replanning. We included the difference due to the slopes to the decision boundary and examined whether it is possible to classify areas as requiring ART or not. <h3>Results</h3> The accuracy rate using the test data was 0.66 for SVM. The decision boundary was shown by the following equation, where f(x) is the DSC and x is the number of fractions. f(x) = -0.0087x + 0.814 The p-values of the ART group and the non-implemented group of the slopes of CTV and the parotid glands were 0.005 and 4.195E-17, respectively, and a significant difference was confirmed (p < 0.05). Normal distribution was created from the slopes of each unimplemented group. The correct answer rate increased to 0.75, assuming that the slope larger than 1σ was an additional condition for ART. <h3>Conclusion</h3> We were able to derive the decision boundary f(x) using machine learning. If the DSC shows a value lower than the f(x) calculated for the same number of fractions, ART is indicated. However, in some cases, ART was performed even though it was not numerically necessary. In other cases, patients who did not undergo ART were judged numerically to require ART. These may be due to the DIR not working well because the tumor changes were intracranial, or due to the oncologist's preference not to perform ART despite large morphological changes. In this study, we proposed the need for ART to approximately 75% of patients. The findings suggest that machine learning may assist the operator with decision-making and replanning.