Deep learning-based auto-segmentation of organs at risk in high-dose rate brachytherapy of cervical cancer

Background and purpose Delineation of organs at risk (OARs), such as the bladder, rectum sigmoid, plays an important role in delivery optimal absorbed dose to target owing steep gradient high-dose rate brachytherapy (HDR-BT). In this work, we propose a deep convolutional neural network-based approach for fast reproducible auto-contouring OARs HDR-BT. Materials methods Images 113 patients with locally-advanced cervical cancer were utilized study. We used ResU-Net network architecture, which uses long short skip connections improve feature extraction procedure accuracy segmentation. Seventy-three chosen randomly training, 10 validation, 30 testing. Well established quantitative metrics, Dice similarity coefficient (DSC), Hausdorff distance (HD), average symmetric surface (ASSD), evaluation. Results The DSC values test dataset 95.7 ± 3.7%, 96.6 1.5% 92.2 3.3% rectum, respectively. HD (mm) 4.05 5.17, 1.96 2.19 3.15 2.03 ASSDs 1.04 0.97, 0.45 0.09 0.79 0.25 Conclusion proposed model achieved good agreement between predicted manually defined contours OARs, thus improving reproducibility contouring workflow.