Dosimetric study of patient specific quality assurance using LINAC log files for volumetric modulated arc therapy

• Main Author: Lew, Kah Seng
• Other Authors: -
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Science::Biological sciences::Biophysics; Science::Physics
• Online Access: https://hdl.handle.net/10356/139756
• Institution: Nanyang Technological University

In radiation therapy, verification of treatment plan is an important step before treatment delivery. This is normally done by medical physicist with a measurement device using gamma analysis as a verification metrics. The measured dose distribution under certain gamma criteria are compared with planned dose distribution to determine the pass/fail of a treatment plan. In this study, the feasibility of using machine learning algorithms on machine trajectory log files obtained during treatment delivery as a verification model was explored. A total of 489 treatment fields consisting of three treatment sites, prostate, spine and thorax was subjected to two different machine learning algorithm, multiple linear regression and one-class support vector machine. These treatment fields are further categorized in training and testing data for machine learning use. Using a gamma criterion of 1%/1mm, results from multiple linear regression using test data shows a mean different of 3.10 ± 4.24 (%) between actual and predicted gamma passing percentage for all treatment site model while prostate, spine and thorax treatment site model shows a mean different of 3.35 ± 3.72 (%), 0.54 ± 0.36 (%) and 2.08 ± 1.55 (%) respectively. On the other hand, one-class support vector machine using a passing percentage of 93% and gamma criterion of 1%/1mm is able to produce a specificity of 63.2% for all treatment sites combined. The respective treatment sites then have a specificity of 76.9% for prostate and 100% for thorax. Specificity refers to the ability of the model to capture false positive. By comparing with our training data log file, a specificity of 93.2% for all treatment sites, 88.0% for prostate treatment site and 100% for thorax treatment site, our test data is within measurement bound and individual model for separate treatment sites are seen to be more accurate for use. The model is also capable of outlier detection due to the nature of the classification algorithm used as compared to multiple linear regression. In this project, one-class support vector machine has the potential to act as a verification metrics for patient specific quality assurance. This will reduce the workload on medical physicist and allow more time for treatment planning.