Abstract / Summary
Abstract Background: Accurate delineation of clinical target volume (CTV) and planning target volume (PTV) is essential for radiotherapy in uterine malignancies. However, the limited soft-tissue contrast of non-contrast computed tomography (NCCT) makes automated target segmentation challenging. Methods: This single-center retrospective study included 157 unique patients. Of these, 137 patients were used for model development and internal testing, including a development cohort of 110 patients and an independent test cohort of 27 patients. A separate cohort of 20 patients was used for workflow evaluation. A multi-label nnU-Net model was developed to simultaneously segment the CTV and PTV. Four nnU-Net configurations were compared using five-fold cross-validation. Segmentation performance was evaluated using the Dice similarity coefficient (DSC) and the 95th percentile Hausdorff distance (HD95). Three radiation oncologists independently assessed the clinical acceptability of the auto-contours using a four-point Likert scale. Workflow efficiency was evaluated by comparing manual contouring with AI-assisted contour editing. Results: The 3D-fullres configuration achieved the numerically highest overall cross-validation performance and was selected for independent internal testing. In the 27-patient independent test cohort, the mean DSC was 90.02% ± 2.33% for PTV and 72.16% ± 4.54% for CTV. The corresponding median HD95 values were 5.12 mm (interquartile range [IQR], 5.00–7.60 mm) and 5.55 mm (IQR, 5.00–9.50 mm), respectively. Rating-level clinical acceptability was 91.4% for PTV and 75.3% for CTV. Inter-rater agreement was higher for PTV than for CTV. In the separate workflow cohort, AI-assisted contouring reduced the mean total task time from 65.4 to 14.2 minutes per case, corresponding to an observed reduction of 78.3% (P<0.001). Conclusion: The multi-label 3D-fullres nnU-Net achieved high PTV overlap and moderate CTV overlap on internal NCCT data and reduced contouring time in a separate workflow cohort. The model may provide editable initial contours, but CTV predictions require careful physician review. External validation and dosimetric assessment are required before routine clinical use.