3223 - Deep Learning Auto-Segmentation of the Internal Pudendal Artery on CT: Dosimetric Validation Against MRI-Informed Ground Truth Contours
Presenter(s)
O. Awad1,2, C. G. Candano1, S. Liang1, Y. Han1, H. Mekdash1, B. Sun1, A. S. Mohamed1,3, D. A. Hamstra1, and Z. A. Siddiqui1; 1Department of Radiation Oncology, Dan L. Duncan Comprehensive Cancer Center, Baylor College of Medicine, Houston, TX, 2Clinical Oncology Department, Faculty of Medicine, Alexandria University, Alexandria, Egypt, 3Department of Radiation Oncology, The University of Texas M.D. Anderson Cancer Center, Houston, TX
Purpose/Objective(s): The internal pudendal artery (IPA) has emerged as a dosimetrically relevant organ at risk for sexual function preservation during prostate radiation therapy. However, CT-based IPA delineation is limited by poor soft tissue contrast, substantial interobserver variability, and inconsistent MRI availability for anatomical guidance. We developed and evaluated a CT-based deep learning (DL) auto-segmentation model for bilateral IPA and assessed whether auto-segmented contours yield dosimetrically equivalent DVH parameters compared to MRI-informed ground truth (GT).
Materials/Methods: A total of 171 planning CTs from prostate cancer patients were included. GT bilateral IPA contours were delineated on CT using CT-MRI registration for anatomical guidance by two attending radiation oncologists. The IPA was defined from the ischial spine anteriorly along the ischioanal fossa to the perineal membrane. A 3D full-resolution nnU-Net v2 model was trained on 140 cases and evaluated on 31 validation cases using Dice similarity coefficient (DSC) and 95th percentile Hausdorff distance (HD95). Dosimetric validation compared DVH parameters (V0-V86) between GT and auto-segmented combined IPA contours using paired Wilcoxon tests, concordance correlation coefficients (CCC), and per-patient mean absolute error (MAE). Clinical threshold concordance was assessed at the previously identified caudal IPA V22 = 93% sexual QOL predictor DVH parameter.
Results: The model achieved a mean DSC of 0.66 ± 0.09 on training and 0.59 ± 0.11 on independent validation, with HD95 of 5.1 ± 3.4 mm and 6.2 ± 3.4 mm, respectively. Despite a moderate geometric agreement, dosimetric concordance was high. In the training cohort (n=140), median per-patient Pearson r was 0.998 with 100% achieving r above 0.95, mean MAE of 3.2%, and V22 of 93% threshold sensitivity of 99%. These results were preserved on independent validation (n=31): median r of 0.996 with 94% achieving r above 0.95, mean MAE of 4.4% with 81% below 5%, and threshold sensitivity of 95% with only 1 of 20 at-risk patients misclassified. No significant differences in dosimetric performance were observed between cohorts (MAE p=0.31, correlation p=0.13).
Conclusion: Despite moderate geometric overlap typical of small tubular vascular structures, CT-based DL auto-segmentation of the IPA produces dosimetrically reliable DVH parameters confirmed on independent validation (r above 0.95 in 94%, MAE below 5% in 81%). The 95% sensitivity at the V22 of 93% threshold ensures nearly all patients at risk for radiation-induced sexual dysfunction are correctly identified. These findings support the clinical feasibility of automated IPA contouring for dose-sparing optimization in prostate radiation therapy planning.