Author: Kellin M De Jesus, Leon Dunn, Les Sztandera, David H. Thomas 👨🔬
Affiliation: IsoAnalytics Pty. Ltd., Thomas Jefferson University 🌍
Purpose: Machine accuracy and performance are critical for ensuring the safety and efficacy of intensity-modulated radiotherapy (IMRT and VMAT). This study aims to analyze a large and diverse set of logfiles from 24 Varian treatment machines (TrueBeam, Edge, Halcyon) to identify statistically relevant performance characteristics. The machines were situated across multiple clinical sites and two countries to identify variations that could impact treatment quality.
Methods: Over 150,000 log files comprised of nine US and 15 Australian machines across multiple clinics were retrospectively analyzed. Logfile processing was conducted using Sentinel, an automated logfile analysis software. Sentinel creates actual and expected images of fluence maps from high-frequency log-file data and performs gamma analysis and other assessments. This dataset was statistically analyzed in terms of gamma passing rates (GPR% 1.5%/0.5mm dose difference/distance to agreement) across individual treatment machines and treatment anatomical sites using IBM SPSS and RStudio to determine whether there are any insights to be gleaned from this large dataset.
Results: Logfiles provide a free, additional source of machine performance information that can be analyzed via software. Analysis of GPRs across multiple machines shows reliable machine performance across multiple treatment sites and machines, with the average gamma pass-rate above 95% at 1.5%/0.5 mm gamma analysis. However, Intra-machine performance variations exist that can be highlighted with modern statistical comparisons like heat maps and pair-wise node comparisons.
Conclusion: Logfile analysis across multiple institutions can provide valuable insights into machine performance, potentially reducing the burden of IMRT/VMAT patient-specific QA by providing complementary information. Integrating image- and data-based machine learning with automated tools offers a scalable approach to building predictive performance models using large datasets, enhancing machine quality assurance processes.