Deep Learning-Based Pelvic Vessel Auto-Segmentation for Standardized Lymph Node Delineation in Prostate Cancer Radiotherapy.
Authors
Affiliations (6)
Affiliations (6)
- Department of Radiation Oncology, University of North Carolina at Chapel Hill, Chapel Hill, NC, USA. Electronic address: [email protected].
- Department of Math and Computer Science, Fayetteville State University, Fayetteville, NC, USA.
- Department of Radiation Oncology, University of North Carolina at Chapel Hill, Chapel Hill, NC, USA.
- Department of Radiation Oncology, University of North Carolina at Chapel Hill, Chapel Hill, NC, USA; Department of Radiation Medicine & Applied Sciences, University of California, San Diego, San Diego, CA, USA.
- School of Information and Library Science, University of North Carolina at Chapel Hill, Chapel Hill, NC, USA.
- Department of Radiation Oncology, University of North Carolina at Chapel Hill, Chapel Hill, NC, USA; School of Information and Library Science, University of North Carolina at Chapel Hill, Chapel Hill, NC, USA; Carolina Health Informatics Program, University of North Carolina at Chapel Hill, Chapel Hill, NC, USA. Electronic address: [email protected].
Abstract
Background Accurate delineation of lymph node clinical target volumes (CTVs) is essential for effective and safe prostate cancer radiotherapy, yet inter-physician variability in pelvic nodal contouring remains high. Because pelvic lymph nodes follow predictable distributions around major pelvic vessels, accurate pelvic vessel segmentation could serve as an anatomic guide to standardize nodal CTV definition. This study aimed to develop an automated pelvic vessel segmentation framework on CT imaging to serve as an anatomical scaffold for lymph node delineation in prostate radiotherapy planning. Methods This single-institution study included non-contrast pelvic CT simulation images from 50 patients with prostate cancer who underwent pelvic nodal irradiation. Major pelvic vessels, including the common, external and internal iliac arteries and veins (including branches of the internal iliac vessels up until they exit the pelvis via the greater sciatic foramen), were manually annotated and checked by expert radiation oncologists and used as reference contours. The dataset was randomly divided into a training cohort (n = 30) and a withheld internal test cohort (n = 20). Automated segmentations were generated using a deep-learning framework trained on physician-annotated CT images. Performance was assessed using the Dice Similarity Coefficient, 95th-percentile Hausdorff Distance and Median Surface Distance, with additional expert review of the automated contours by an experienced radiation oncologist. Results On the withheld internal test cohort, the automated segmentation performance achieved a mean (95% CI) Dice Similarity Coefficient of 0.92 (0.91-0.92), Median Surface Distance of 0.70 mm (0.55-0.86) and 95th-percentile Hausdorff Distance of 9.88 mm (8.43-11.33). Conclusions and Relevance Automated pelvic vessel segmentation on non-contrast CT demonstrates quantitative agreement with physician contours. By providing a patient-specific vascular anatomy reference, this approach may provide a foundation for future anatomy-guided nodal CTV delineation workflows, pending external validation and prospective evaluation of clinical impact.