S122
Brachytherapy - Physics
ESTRO 2026
Keywords: Brachytherapy, Monte Carlo, dosimetry References: [1] Famulari G et al. RapidBrachyMCTPS: a Monte Carlo-based treatment planning system for brachytherapy applications. Physics in Medicine & Biology. 2018 Aug;63(17):175007.[2] Glickman H et al. RapidBrachyMCTPS 2.0: a comprehensive and flexible Monte Carlo-based treatment planning system for brachytherapy applications. arXiv:2007.02902. 2020 Jul.[3] Beaulieu L et al. AAPM WGDCAB Report 372: a joint AAPM, ESTRO, ABG, and ABS report on commissioning of model - based dose calculation algorithms in brachytherapy. Medical physics. 2023 Aug;50(8):e946-60.[4] Ma Y et al. A generic TG - 186 shielded applicator for commissioning model - based dose calculation algorithms for high - dose - rate 192Ir brachytherapy. Medical physics. 2017 Nov;44(11):5961-76. Validation of an auto-segmentation solution for the delineation of regions of interest in image guided brachytherapy treatment of cervix cancers Ceri Doherty 1 , Abdulkerim Duman 2 , Robert Chuter 3,4 , Michael Hutton 5 , Emiliano Spezi 2 1 Medical Physics, Velindre University NHS Trust, Cardiff, United Kingdom. 2 School of Engineering, Cardiff University, Cardiff, United Kingdom. 3 Christie Medical Physics and Engineering, The Christie NHS Foundation Trust, Manchester, United Kingdom. Digital Poster Highlight 1410 4 Division of Cancer Sciences, School of Medical Sciences, Faculty of Biology, Medicine and Health, University of Manchester, Manchester, United Kingdom. 5 INIR/MR Section, Newcastle Upon Tyne NHS Foundation Trust, Newcastle Upon Tyne, United Kingdom Purpose/Objective: To train and evaluate the performance of deep learning auto-segmentation for the delineation of ROIs on MRI for cervical cancer Image Guided Brachytherapy (IGBT) due to a lack of commercial alternatives. To validate auto-segmentation methods against inter-observer variability of manual contouring and associated dose parameters, alongside manual contouring timing data. Material/Methods: A ‘gold-standard dataset’ of 104 anonymised MRIs using the Venezia applicator was compiled on Oncentra Brachy (Elekta AB, Sweden). Three established algorithms, U-Net, SegNet and PSPNet, were trained to segment the Cervix, GTV, Bladder, Bowel and Rectum [1,2,3]. Performance was evaluated using quantitative metrics (Dice Similarity Coefficient [DSC], Hausdorff Distance [HD], and mean Distance to
Agreement [mDTA]) and qualitative clinical observer scoring.Manual contouring inter-observer variability was evaluated using 3 previous clinical cases, with 3 oncologists delineating target volumes and 11 treatment planners contouring OARs. Variation was assessed using DSC, HD and dose parameters from the clinical treatment plan. A timing study of 43 cases was performed during clinic to establish manual contouring time. Dosimetric evaluation was completed for the auto-segmentation models with the clinical treatment plan applied to 10 patients, alongside a small study into the time required to manually adapt OAR contours until suitable for clinical use. Results: The U-Net auto-segmentation model was quantitatively evaluated to produce contours closest to the clinical volumes (Figure 1). The most acceptable results were seen for the Bladder, with Rectum being the poorest. The qualitative scoring established the U- Net model produced contours of moderate clinical use for all ROIs other than Rectum (Figure 2). All PSPNet contours were qualitatively assessed to be clinically unsuitable.
Improved DSC and HD results were established for the inter-observer variability study compared to the auto- segmentation methods. However, dosimetric evaluation variations for both the inter-variability study and auto-segmentation models were found to be comparable to each other and those reported in the literature [4]. Mean manual contouring time was 54.6 minutes, of which 33.3 minutes was OAR delineation. Manual adaption of U-Net produced OARs was recorded to take 7.1 minutes. Alongside the estimated auto-segmentation workflow this projects
Made with FlippingBook - Share PDF online