Purpose: Developing automated methods to identify task-driven quality assurance (QA) procedures is key toward increasing safety, efficacy, and efficiency. We investigate the use of machine learning (ML) methods for possible visualization, automation, and targeting of QA, and assess its performance using multi-institutional data.
Methods: To enable automated analysis of QA data given its higher dimensional nature, we used nonlinear kernel mapping with support vector data description (SVDD) driven approaches. Instead of using labeled data as in typical support vector machine (SVM) applications, which requires exhaustive annotation, we applied a clustering extension of SVDD, which identifies the minimal enclosing hypersphere in the feature space defined by a kernel function separating normal operations from possible failures (i.e., outliers). In our case, QA test data are mapped by a Gaussian kernel to a higher dimensional feature space and then the minimal enclosing sphere was identified. This sphere, when mapped back to the input data space along the principal components, can separate the data into several components, each enclosing a separate cluster of QA points that could be used to evaluate tolerance boundaries and test reliability. We evaluated this approach for gantry sag, radiation field shift, and [multileaf collimator (MLC)] offset data acquired using electronic portal imaging devices (EPID), as representative examples.
Results: Data from eight LINACS and seven institutions (n = 119) were collected. A standardized EPID image of a phantom with fiducials provided deviation estimates between the radiation field and phantom center at four cardinal gantry angles. Deviation measurements in the horizontal direction (0°, 180°) were used to determine the gantry sag and deviations in the vertical direction (90°, 270°) were used to determine the field shift. These measurements were fed into the SVDD clustering algorithm with varying hypersphere radii (Gaussian widths). For gantry sag analysis, two clusters were identified one of which contained 2.5% of the outliers and also exceeded the 1 mm tolerance set by TG-142. In the case of field shifts, SVM clustering identified two distinct classes of measurements primarily driven by variations in the second principal component at 270°. Results from MLC analysis identified one outlier cluster (0.34%) along Leaf offset Constancy (LoC) axis that coincided with TG-142 limits.
Conclusion: Machine learning methods based on SVDD clustering are promising for developing automated QA tools and providing insights into their reliability and reproducibility.
Keywords: SVM; Linacs; higher dimension visualization; machine learning; quality assurance.
© 2019 American Association of Physicists in Medicine.