PO0104

• Published in: Brachytherapy Vol. 23; no. 6; pp. S87 - S88
• Main Authors: Amjad, Asma, PhD, Genz, Diana, BS, Iandolo, Riccardo, PhD, Balia, Maddalena, PhD, Morrow, Natalya, PhD, Prior, Phillip, PhD, Paulson, Eric, PhD, Bedi, Meena, MD, Rownd, Jason, MS, Erickson, Beth, MD
• Format: Journal Article
• Language: English
• Published: 01.11.2024
• Subjects: Hematology, Oncology, and Palliative Medicine; Radiology
• ISSN: 1538-4721
• DOI: 10.1016/j.brachy.2024.08.126

PurposeHigh dose rate (HDR) GYN Brachytherapy (BT) is a time sensitive adjuvant definitive radiation treatment for endometrial cancer following surgical staging. Currently volumetric imaging, with contours of organs at risk (bladder and rectum, bowels depending on proximity to applicator) and HR-CTV are used for online treatment planning. On average 20-40 minutes are expected for contouring, subject to anatomical complexity, image registration and uninterrupted availability of Physicians. Deep learning based automatic segmentation (DLAS) is expected to improve efficacy of such workflows, especially for in-patient or sensitive care patients. Materials and MethodsIn this study a novel three-level deep learning architecture consisting of: i) a recursive neural network for organ detection, ii) convolutional neural networks for probability maps generation and ii) multi-cephalic ensemble neural network to introduce anatomical consistency though the segmentation by deformation principle, was used to segment bladder and rectum. The organs were chosen due to the high variability in their shapes attributed to large and arbitrary deformations e.g., bladder filling, and large air pockets in the rectum. While the developed tool was trained on CT images with a vaginal cylinder applicator in place, we tested the tool for 20 vaginal cylinder and 5 Syed obturator cases with applicators in place. The segmentation quality and accuracy were quantified using visual and well-established 3D metric analysis, including Dice Similarity Coefficient (DSC), Mean Distance to Agreement (MDA) and 95 percentile Hausdorff Distance (HD95). Furthermore, to remove clinical contouring preference bias, a rectum contour matching the length of the clinical contour was derived from the auto-segmented rectum. Additionally, dosimetric parameters including D 0.1cc, D 1cc and D 2cc of DLAS and clinical contours were also calculated and compared for selective cases using the original BT clinical plans. ResultsAverage 3D accuracy metrics DSC, MDA/mm, HD95/mm values for cylinder cases were (0.88, 1.87. 6.32), (0.68, 4.84, 20.1), (0.86, 1.6, 15.3) and for Syed cases (0.7, 5.76, 18.7) and (0.55, 7.2, 22.6) for bladder, rectum, and clinical-match rectum, respectively. The figure below shows a scatter plot of DSC vs MDA, each symbol represents a test case. The inset show a typical box plot of the 3D DSC for the cylinder cases only. As expected, rectum accuracy improved, when the DLAS was reduced to match clinical contouring guidelines. The percent difference in the dose parameters comparison was found to be primarily under 10%, with a few cases above 20%. Poor segmentation accuracy was observed for Syed obturator cases. This was attributed to i) absence of training data during algorithm development, ii) to presence of needles, iii) contrast in rectum and iv) the Foley catheter. ConclusionsIn this study we show the potential of using DL-based segmentation tools for GYN HDR Brachytherapy clinical workflows. This approach is expected to help streamline the contouring process and remove operator bias.