146 - Cross-Institutional Validation of a Mixed-Cohort-Trained 3D Res-U-Net for Cervical-Bed CTV Segmentation in Cervical Cancer Radiotherapy
Presenter(s)
W. C. You1,2, C. Y. Lai3, Y. T. Wu3, K. L. Yang4, K. H. Yao5, C. W. Jao3, C. Y. Lin3, Y. F. Lu6, Y. Y. Hsu7, M. S. Chi4, C. C. Wen4, and C. H. Hsu4; 1Department of Radiation Oncology, Taichung Veterans General Hospital, Taichung, Taiwan, 2Department of Post-Baccalaureate Medicine, National Chung Hsing University, Taichug City, Taiwan, 3Institute of Biophotonics, National Yang Ming Chiao Tung University, Taipei City, Taiwan, 4Department of Radiation Therapy and Oncology, Shin Kong Wu Ho-Su Memorial Hospital, Taipei, Taiwan, 5Institute of Biophotonics, National Yang Ming Chiao Tung University, Taipei, Taiwan, 6Department of Radiation Oncology, Taichung Veterans General Hospital, Taichung, Taiwan, Taichung City, Taiwan, 7Department of Radiation Oncology, Taichung Veterans General Hospital, Taichung City, Taiwan
Purpose/Objective(s):
Precise delineation of the clinical target volume (CTV) is critical for maximizing tumor control while minimizing toxicity in cervical cancer radiotherapy. However, automated segmentation of postoperative cervical-bed CTV remains challenging because of indistinct anatomical boundaries on CT. This study aimed to develop and validate a deep learning–based auto-segmentation framework and to determine whether incorporating anatomically related cohorts during training enhances localization accuracy, robustness, and cross-institutional generalizability, thereby supporting both offline planning and online adaptive radiotherapy workflows.Materials/Methods: This retrospective study included CT scans from 141 cervical cancer patients treated between 2020 and 2023, comprising 114 CTV–cervical and 28 CTV–cervical bed (post-surgical) cases. The cohort was randomly split into training and internal test sets at an 8:2 ratio, and an external test set of 10 cases was collected from an independent institution. We compared a unified training strategy (mixed-cohort) against group-specific models. A two-stage 3D Res-U-Net pipeline was adopted: stage 1 detected the lowest point of the left femoral head to define a pelvic-cavity VOI and cropped the CT volume from 512×512×~200 to 280×440×65 voxels; stage 2 performed CTV segmentation within the VOI. Models were trained for 500 epochs using Adam (lr=) with Dice loss (DL). Performance was evaluated using DSC, recall, precision, and HD95.
Results: For CTV–cervical segmentation, performance was comparable across training strategies in the internal cohort (DSC: 0.85 ± 0.06 for both), although cervical-only training showed slightly improved boundary accuracy (HD95: 2.77 ± 2.94 mm vs 3.68 ± 5.72 mm). Similar robustness was observed in the external cohort (DSC: 0.77 ± 0.05–0.06; HD95: 3.12 ± 0.84–3.36 ± 0.74 mm). In contrast, for the more anatomically ambiguous cervical-bed target, mixed-cohort training markedly outperformed cervical-bed–only training, achieving higher internal DSC (0.75 ± 0.04 vs 0.64 ± 0.08) and lower HD95 (3.74 ± 2.26 mm vs 7.78 ± 3.97 mm). These gains persisted in external validation (DSC: 0.63 ± 0.05 vs 0.46 ± 0.05; HD95: 6.16 ± 0.91 mm vs 10.21 ± 2.70 mm), with a notable improvement in recall (0.88 ± 0.05 vs 0.54 ± 0.17), indicating enhanced target coverage and segmentation stability.
Conclusion: Our study demonstrates that joint training on CTV–cervical and CTV–cervical bed consistently improves cervical-bed segmentation, as evidenced by higher DSC and substantially lower HD95 in both internal and external evaluations. These gains suggest that transferable anatomical priors learned from cervical cases enhance cervical-bed localization and boundary consistency. Although inter-institutional variability remains a factor in external datasets, mixed-cohort training offers a robust strategy to overcome data scarcity and improve the generalizability of auto-contouring models in clinical practice.