Using a Deep Learning System to Delineate the Elective Pelvic Nodal Volumes in Patients with Prostate Cancer: A Tool Aimed to Facilitate Peer Review

Background/Objective The delineation of elective pelvic lymph node-bearing tissues in patients with prostate cancer can be challenging, and such segmentations have been associated with significant inter-physician variations and inaccuracies. Peer-review of physician-defined image segmentations is a common approach aimed to mitigate these types of errors. However, peer review is labor intensive and is not readily available in many settings (e.g. smaller and rural facilities). We therefore created, and herein assess the clinical appropriateness of, an artificial intelligence (AI) based tool to delineate the elective pelvic nodal volumes in patients with prostate cancer. Methods This study was conducted at a single institution and involved the treatment planning computed tomography scans of 216 patients with prostate cancer who received elective pelvic nodal irradiation from 2021 to 2024. Patient images were split into an 80:20 training and internal validation set, resulting in 179 training and 41 validation images. The completely withheld test set was performed on 10 completely withheld prostate nodal irradiation cases. Data were analyzed from March 2025 to June 2025. Segmentation performance was quantified as concordance with expert-defined contours measured using the Dice Similarity Coefficient, Median Surface Distance, and Hausdorff Distance. Results Segmentation performance in the completely withheld test data set achieved a median (95% CI) Dice similarity coefficient of 0.79 (0.71-0.86), median surface distance of 2.07mm (0.83-3.31mm), and Hausdorff Distance of 27.14mm (15.85-38.08mm). Conclusions and Relevance These results suggest that AI-based segmentations tool can provide a reasonable standard reference to promote meaningful discussion during peer review in radiation oncology.