ESTRO 2021 Abstract Book

S768

ESTRO 2021

Conclusion The assessment of auto-contouring performance in clinical routine has highlighted good geometrical agreement, but still user adjustments are needed in the majority of cases. The lungs and spiral cord showed best geometrical agreement, with no need to change these contour unless big errors are found. Furthermore, a significant time-saving was confirmed for the LC group. PD-0927 MRI-based deep learning auto-contouring for organs-at-risk in gynecological brachytherapy P. Gonzalez 1 , A. Mans 1 , E. Schaake 1 , M. Nowee 1 , U. van der Heide 2 , R. Simões 2 1 Antoni van Leeuwenhoek - Netherlands Cancer Institute, Department of Radiation Oncology, Amsterdam, The Netherlands; 2 Antoni van Leeuwenhoek - Netherlands Cancer Institute, Department of Radiation Oncology, Amsterdam, The Netherlands Purpose or Objective Manual contouring of organs-at-risk (OAR) as part of gynecological brachytherapy (BT) treatment planning is a time-consuming process. Between applicator placement and irradiation the patient has to endure the discomfort of the applicator while waiting. Deep learning based auto-contouring is expected to help improve both contour quality, reduce the clinical workload and more importantly reduce waiting times and thereby reduce patient discomfort. The aim of this work is to assess the geometrical and dosimetric performance of a state-of-the-art deep-learning based segmentation framework for the organs-at-risk of gynecological BT on MR images. Materials and Methods 194 gynecological cancer patients treated with interstitial and intracavitary brachytherapy between 2011 and 2019 were selected from the clinical database with a total of 525 separate fractions. For each patient the bladder, rectum, sigmoid and small bowel were delineated by an experienced clinician on T2 weighted MRI scans. The dataset was split into 70% of the patients for training, 15% for testing and 15% for validation. The patients were stratified according to applicator type, namely cylinder-type and intra-uterine applicators. The nnUnet , a state-of-the-art deep learning-based segmentation method that automatically configures itself, was trained and used to automatically segment the structures. The Dice similarity coefficient (DSC), the mean surface distance (MSD) and the 95 th percentile Hausdorff distance (95HD) were used to assess the geometrical similarity between automatic and manual contours. Furthermore, to evaluate the dosimetric performance of the network, the D2cc was calculated from the dose distribution on both the automatic and manual structures. Results Figure 1 shows the results of the geometrical and dosimetric performance metrics on the test set. For all three geometrical metrics the nnUnet performed best for the bladder. The 95HD is largest for the sigmoid, 3.1 cm with an inter-quartile range of 3.0 cm, whilst the Dice score (0.50) is lowest for the small bowel. For the rectum, sigmoid and small bowel, the errors are largest in the most caudal and cranial areas. The dosimetric performance was best for the bladder with an average ΔD2cc of (-0.1 ± 0.3) Gy and worst for the small bowel with an average ΔD2cc of (-0.3 ± 0.9) Gy. Figure 2 shows an example of a 3D rendering of all four organs together with the D2cc isosurface. The trained network took on average 15 seconds to contour one patient on a Nvidia GRID T4-16Q virtual GPU.