3006 - Deep Learning-Based Pelvic Vessel Auto-Segmentation for Standardized Lymph Node Delineation In Prostate Cancer Radiotherapy
Presenter(s)
Y. Gao1, D. Wells2, S. Bhattacharya3, M. C. Repka4, E. Lavrova5, B. M. Anderson4,6, S. K. Das7, A. Rajasekar8, E. Byrd1, R. McGurk4, J. Dooley7, K. Adapa9, J. L. Wright7, L. B. Marks4, and L. Mazur4,10; 1UNC Chapel Hill, Chapel Hill, NC, 2Fayetteville State University, Fayetteville, NC, 3UNC, Chapel Hill, NC, 4Department of Radiation Oncology, University of North Carolina, Chapel Hill, NC, 5University of North Carolina, Department of Radiation Oncology, Chapel Hill, NC, 6University of California San Diego, Department of Radiation Medicine and Applied Sciences, La Jolla, CA, 7Department of Radiation Oncology, University of North Carolina School of Medicine, Chapel Hill, NC, 8University of North Carolina at Chapel Hill, Chapel Hill, NC, 9Department of Radiation Oncology, University of North Carolina at Chapel Hill, Chapel Hill, NC, 10Carolina Health Informatics Program, School of Information and Library Science, University of North Carolina, Chapel Hill, NC
Purpose/Objective(s):
Elective pelvic lymph node clinical target volumes (CTVs) in prostate radiotherapy are defined as margin expansions around the iliac vessels; however, pelvic nodal contouring remains highly variable across physicians and institutions, even when consensus atlases are applied. Much of this variability reflects differences in how perivascular margins are interpreted and implemented in practice. Because lymphatic drainage follows predictable patterns along vascular anatomy, we hypothesized that automated segmentation of pelvic vessels on planning CT could achieve geometric agreement comparable to physician-generated contours. This framework was designed to provide an objective, anatomy-based scaffold to support future efforts toward more standardized nodal CTV delineation and peer-review workflows.Materials/Methods:
Non-contrast pelvic CT simulation scans from 50 patients undergoing pelvic nodal irradiation were analyzed. Major pelvic vessels, including the common, external and internal iliac arteries and veins with distal branches, were manually contoured by two physicians (contoured by a resident and subsequently reviewed by an experienced radiation oncology faculty). Data were randomly divided into training (n=30) and independent testing (n=20) cohorts. A 3D full-resolution nnU-Net framework was trained using a combined Dice and cross-entropy loss with five-fold cross-validation. Standard nnU-Net data augmentation (random rotations, mirroring, and Gaussian noise) was applied during training. Performance on the completely withheld test cohort was evaluated using Dice similarity coefficient (DSC), median surface distance (MSD), and 95th percentile Hausdorff distance (HD95), quantifying volumetric overlap and boundary agreement, as well as subjective review.
Results:
On the independent test cohort, automated segmentations achieved a median (95% CI) DSC of 0.92 (0.91-0.92), MSD of 0.70 mm (0.55-0.86), and HD95 of 9.88 mm (8.43-11.33). The high DSC together with sub-millimeter MSD indicates strong volumetric overlap and geometric fidelity between predicted and physician-contoured vessel contours. The boundary deviations were primarily observed in small-caliber distal branches.Conclusion:
Automated pelvic vessel segmentation on non-contrast planning CT demonstrates high quantitative agreement with physician-generated contours. Because nodal CTVs are defined relative to vascular anatomy, this approach provides a reproducible, patient-specific structural reference that may facilitate more consistent nodal CTV delineation across clinicians and institutions. This framework likely also be equally applied to other disease sites in which nodal volumes are defined relative to vascular anatomy. Ongoing work includes expansion to larger multi-institutional datasets and prospective evaluation of its impact on inter-observer variability in nodal CTV delineation.