CT-based deep learning auto-segmentation of high-risk clinical target volume in CT-guided cervical cancer
Chen-Ying Ma1, Yi Fu1, Lu Zhang1
1Department of Radiation Oncology, The First Affiliated Hospital of Soochow University, Suzhou, China.
Introduction:
Computed tomography (CT)-based high-risk clinical target volume (HR-CTV) auto-segmentation has been previously investigated, but evidence remains heterogeneous across applicators, target definitions, architectures, and clinical evaluation procedures. A pragmatic within-cohort benchmark of 2D U-Net, 3D U-Net, and nnFormer was performed in a CT-only, applicator-in-situ workflow.
Methods:
CT images from 544 brachytherapy fractions in 182 patients were analyzed, including 509 fractions from 163 patients treated with tandem-and-ovoid applicators and 35 fractions from 19 patients treated with vaginal cylinders. All fractions from a patient remained in one partition (HR-CTVc: 325/82/102 fractions; HR-CTVv: 22/6/7 fractions). Only seven test fractions were available and thus the postoperative HR-CTVv analysis was exploratory. Performance was assessed using Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (HD95), average surface distance (ASD), and physician consensus edit categories. Task-level CT normalization parameters were estimated exclusively from the training partition and fixed for validation and testing; spatial cropping was image centered and contour independent.
Results:
In the intact-cervix test cohort, 3D U-Net showed the most favorable descriptive combination of overlap and surface agreement (DSC 0.848 ± 0.059; HD95 2.755 ± 1.688 mm; ASD 1.008 ± 0.622 mm). In the exploratory vaginal stump cohort, corresponding values were 0.788 ± 0.060, 4.724 ± 2.434 mm, and 2.917 ± 2.150 mm. On physician consensus review, 91/102 HR-CTVc and 6/7 HR-CTVv 3D U-Net contours required no or localized correction. In an ancillary 20-fraction independent-contouring analysis, physician-to-physician HR-CTV agreement was DSC 0.80 ± 0.06 and HD95 3.4 ± 1.8 mm.
Conclusions:
In this internally validated CT-guided cohort, 3D U-Net provided the most favorable overall performance among the evaluated architectures for intact-cervix cases. The postoperative results provide a preliminary feasibility signal, and confirmation in a larger cohort is required. These findings support further physician-supervised implementation, followed by multicenter external validation and cohort-level dosimetric assessment.
