3411 - Construction and Selection of the Automatic Segmentation Tools for High-Risk Clinical Tumor Volume in Brachytherapy for Cervical Cancer Based on Deep Learning

C. Ma¹, L. Zhang¹, Y. Wang², L. C. Jia³, W. Q. Xiong², W. Zhang², X. Xu¹, and J. Zhou¹; ¹the First Affiliated Hospital of Soochow University, Suzhou, China, ²Shanghai United Imaging Healthcare Co., Ltd., Shanghai, China, ³Shenzhen United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, China

Purpose/Objective(s): To verify the clinical applicability of automatic segmentation tools of high-risk clinical tumor volume (HR-CTV) based on 2D-Unet, 3D-Unet and nnformer networks with two different applicators in brachytherapy for cervical cancer. Materials/

Methods: Computed Tomography (CT) images of 544 fractions of cervical cancer brachytherapy from 182 patients were selected, including 163 cases (509 fractions in total) of locally advanced patients using standard III tube applicators, and 19 (35 fractions in total) of postoperative patients using vaginal applicators. Cervical HR-CTV_c and vaginal stump HR-CTV_v were contoured respectively according to the HR-CTV definitions of CT image-guided brachytherapy for cervical cancer patients in GEC ESTRO and the American Brachytherapy Society. It was divided into training set, verification set and testing set according to the ratio of 325:82:102 and 22:6:7 respectively. The accuracy of the prediction model was evaluated using dice similarity coefficient (DSC), 95% Hausdorff distance (95HD) and average surface distance (ASD). At the same time, the automatic segmentation results of the rectum, bladder, sigmoid colon, and small intestine were reviewed by senior physicians.
Results: For HR-CTV_c: the training results of the 3D-Unet network were DSC 0.848 ± 0.059, 95HD 2.755 ± 1.688 mm and ASD 1.008 ± 0.622 mm; the training results of the 2D-Unet network were DSC 0.817 ± 0.061, 95HD 3.588 ± 2.196 mm and the ASD 1.580 ± 1.744 mm; the training results of the nnformer network were DSC 0.806 ± 0.061, 95HD 10.360 ± 68.329 mm and the ASD 2.145 ± 7.981 mm. For HR-CTV_v: the training results of the 3D-Unet network were DSC 0.788 ± 0.060, 95HD 4.724 ± 2.434 mm and the ASD 2.917 ± 2.150 mm; the training results of the 2D-Unet network were DSC 0.766 ± 0.093, 95HD 5.952 ± 3.117 mm and the ASD 2.176 ± 0.890 mm; the training results of the nnformer network were DSC 0.706 ± 0.091, the 95HD 5.034 ± 1.523 mm and the ASD 2.462 ± 0.787 mm. Comparing 2D and 3D-Unet, all parameters of the training results of HR-CTV_c were significantly different (P<0.05), and there was no significant difference in HR-CTV_v. Comparing 2D/3D-Unet with nnformer, in the training results of HR-CTV_c and HR-CTV_v, only the DSC value had a significant difference (P<0.05). In addition, the OAR automatic segmentation model automatically completed the outline of the rectum, bladder, sigmoid colon, and small intestine before the HR-CTV delineation process. After review by senior physicians, the above OAR structures did not need to be manually modified.
Conclusion: The HR-CTV_c and HR-CTV_v with two different applicators in brachytherapy contoured by the automatic segmentation tool based on the 3D-Unet network had the best effect, which compared with the manual outline by senior physicians. The HR-CTV_c and HR-CTV_v could be used in clinical practice after being combined with the rectum, bladder, sigmoid colon, and small intestine to form the outline template respectively.