ESTRO 2025 - Abstract Book

S3227

Physics - Intra-fraction motion management and real-time adaptive radiotherapy

ESTRO 2025

al. Quantifying the reduction of respiratory motion by mechanical ventilation with MRI for radiotherapy. Radiat Oncol. 2022;17(1):99. 2. Veldman JK, Parkes MJ, Stevens MF, van Duren KML, van Kesteren Z, van den Aardweg JG, et al. Rapid non invasive mechanical ventilation appears superior to non-invasive high-frequency jet ventilation in reducing respiratory motion for radiotherapy. JCA Advances. 2024;1(3-4). 3. Zachiu C, Papadakis N, Ries M, Moonen C, Denis de Senneville B. An improved optical flow tracking technique for real-time MR-guided beam therapies in moving organs. Phys Med Biol. 2015;60(23):9003-29.

1452

Digital Poster First implementation of an AI-enabled marker tracking algorithm on a clinical radiation therapy system Chandrima Sengupta 1 , Benjamin Zwan 2 , Freeman Jin 1 , Paul Keall 1 1 Image X Institute, University of Sydney, Sydney, Australia. 2 Department of Radiation Oncology, Central Coast Cancer Centre, Gosford, Australia Purpose/Objective: The patient benefits of real-time three-dimensional (3D) image-guided radiation therapy (IGRT) have been shown for MRI and x-ray guided prostate cancer stereotactic ablative radiation therapy (SABR).[1, 2] One of the challenges associated with performing precise real-time 3D IGRT is segmenting tumours in x-ray images with poor soft tissue contrast along with large and fast motion of tumours. To address this challenge, we advanced an artificial intelligence (AI) algorithm for improved marker detection [3]. The aim of this study was to characterise the performance of the real-time AI-enabled marker tracking algorithm for its clinical use following existing best practice guidelines.[4-6] Material/Methods: A technology to enable real-time 3D IGRT on standard linear accelerators is Kilovoltage Intrafraction Monitoring (KIM), which has been clinically trialled in a multi-institutional setting using a template matching based non-AI segmentation method.[7, 8] To improve KIM’s segmentation performance, a convolutional neural network (CNN) was trained for tracking gold fiducial markers and tested using the Neural Network Toolbox in MATLAB on prostate[7] and liver[8] patient measured data. To enable fast real-time implementation, the CNN model was converted into a dynamic-link library (dll) and added to the C# KIM framework. For real-time performance characterisation of AI-enabled KIM, an anthropomorphic phantom with three gold markers was positioned on a robotic motion platform (Fig. 1), capable of reproducing patient-measured motion traces.[9] The validation tests included static and dynamic tests using patient-measured prostate and liver motion traces.[6] The ground truth was defined as the known shifts from the treatment isocentre for the static tests and patient-measured motion traces programmed into the robot for the dynamic tests. The geometric accuracy and precision were quantified by the mean and standard deviation of the differences between the AI-enabled KIM measured motion with the ground truth motion (Fig. 2). The AI-enabled KIM latency was defined as the measured time difference between the Varian RPM (with known latency [10]) and KIM-measured motion.