2154 - Development of an AI-Assisted Automated Quality Assurance AnalysisTool for Linear Accelerators Using Pylinac
Presenter(s)
Y. W. Lin, Y. W. Hsieh, P. ChiaPeng, C. S. Liu, H. H. Tseng, and Y. H. Yen; Department of Radiation Oncology, Kaohsiung Veterans General Hospital, Kaohsiung, Taiwan
Purpose/Objective(s): This study aims to develop an automated, image-based quality assurance (QA) tool for linear accelerators (Linacs) using the open-source Pylinac library, with AI-assisted programming via ChatGPT-4o. The objective is to enhance the feasibility and automation of QA procedures, enabling resource-limited radiation therapy centers to effectively perform TG-142 QA analysis, ensuring treatment equipment stability and safety.
Materials/Methods: The AI-assisted development process leveraged ChatGPT-4o to generate Python code integrating Pylinac 3.31.0, an open-source software package containing 15 modules. Five modules were tested: CatPhan, ACR Phantom, Starshot, Picket Fence, and Field Analysis, to automate the analysis of CBCT and EPID images from an O-ring Linac. ChatGPT-4o facilitated the generation, execution, and optimization of Python scripts for QA analysis, streamlining the workflow for users without prior programming experience. The automated analysis pipeline enhanced data extraction, processing, and reporting, improving the reproducibility and efficiency of QA evaluations.
Results: The AI-assisted approach enabled users without prior programming experience to efficiently develop QA analysis modules and automatically generate analytical reports using ChatGPT-4o. The automated analysis yielded the following results:For the CatPhan 504 Phantom, the measured CT numbers were -1000 HU for air, -197 HU for PMP, -114 HU for LDPE, -58 HU for polycarbonate, 100 HU for acrylic, 321 HU for Delrin, and 924 HU for Teflon. The high-contrast resolution was 0.41 lp/mm at MTF 50%, while the HU uniformity was -10.8 ± 1.17 HU. The low-contrast detectability test identified six visible ROIs.For the ACR Phantom, the CT numbers measured -1000 HU for air, 105 HU for polyethylene, 108 HU for acrylic, 948 HU for bone, and -11 HU for water, with a low-contrast resolution CNR of 0.68, HU uniformity of -10.2 ± 1.6 HU, and a high-ontrast resolution of 0.55 lp/mm at MTF 50%.The Starshot analysis indicated a minimum circle diameter of 0.20 mm, while the Picket Fence test detected a maximum positional error of 0.114 mm. The field size measurements for a 10×10 cm² field were 100.2 mm in the horizontal direction and 98.8 mm in the vertical direction.
Conclusion: ChatGPT-4o successfully facilitated the development of a Pylinac-based QA tool for Linacs, eliminating the need for prior programming expertise. This AI-assisted approach significantly enhances the efficiency of QA analysis, offering a cost-effective solution for resource-limited radiation therapy centers. The implementation of AI-driven automation in QA procedures has the potential to streamline clinical workflows, ensuring precise and reliable radiation therapy quality control.