S2214
Physics - Intra-fraction motion management and real-time adaptive radiotherapy
ESTRO 2026
system reliably generated consistent anatomical structure overlays, including surrounding organs and tumor projections, across all processed 4D-CT datasets. On a standard GPU configuration (NVIDIA GeForce GTX 1050), the generation of DRR images requires approximately 40 seconds per image pair, considering a single respiratory phase of the CT.Preliminary validation confirms accurate tumor visibility within the DRR projections and consistent bounding box initialization for deep learning–based
developed to automatically generate DRRs from planning CTs and to project anatomical structures segmented with the TotalSegmentator tool onto the DRR coordinate space. This pipeline produces paired image–label datasets suitable for deep learning training (see. figure 1).A two-step AI approach is being developed. First, a YOLO-based object detection model identifies the tumor region in each stereoscopic DRR image in real time. In a second stage, a segmentation module refines the tumor boundaries within the detected region. Additional work will evaluate temporal consistency constraints and uncertainty modelling to enhance robustness against respiratory motion variability.
localization. Conclusion:
A scalable and automated workflow was established to prepare clinical lung imaging datasets for AI-based real-time tumor localization in stereoscopic DRRs. The proposed two-step detection and segmentation framework is expected to support improved intrafraction motion management and has the potential to reduce treatment margins in image- guided adaptive radiotherapy. Further evaluation will assess accuracy, inference latency, and clinical integration feasibility. Keywords: AI, Lung tumor, Real-Time Tracking
Digital Poster Highlight 1146
An investigation of prostate SBRT treatments feasibility without fiducial markers through intrafractional prostate motion analysis Sabrina Côté-Maldonado, Levon Igidbashian, Aimée Lauzon, Cédric Filion, Sabrina Hanchay, Caroline Pagé,
Camille Dupont, Sterlin Bai, Gary Mok Radiation Oncology, CICL, Laval, Canada
Purpose/Objective Ultra-hypofractionated regimens have been widely studied for prostate cancer treatments. The PACE-B clinical trial particularly demonstrated non-inferiority of SBRT compared to conventional radiotherapy regarding biochemical and clinical failure for low or intermediate risk localized prostate cancer [1]. The trial protocol strongly recommended use of fiducial markers for image guidance [2]. Resource constraints, delays to treatment and the invasive nature of fiducials were barriers to implementing the technique at our centre. The purpose of this study was to confirm the feasibility of delivering prostate SBRT treatments without the use of fiducial markers. Material/Methods A 36.25Gy/5fx prostate SBRT treatment technique was implemented at our center. To ensure 5mm isotropic PTV margins were appropriate without the use of fiducials, sources of uncertainty were minimized by improving CT/MRI image resolution, limiting treatment time and ensuring reproducible patient anatomy. The intrafractional prostate motion was estimated for the
Figure 1 : DRRs and structures generated from the 0% phase of the 4D CT. Lung contours (blue), vertebrae (green), heart (orange), and tumor (red) are shown. Results: The automated DRR and contour projection workflow was successfully implemented, enabling large-scale dataset creation without manual intervention. The
Made with FlippingBook - Share PDF online