Author: Brian M. Anderson, Shiva K. Das, Meagan Foster, Anirudh Karunaker, Lawrence B. Marks, Lukasz Mazur, Michael Repka 👨🔬
Affiliation: UNC Chapel HIll, University of North Carolina at Chapel Hill, UNC School of Medicine, University of North Carolina 🌍
Purpose: Development of a peer review segmentation check system to identify deviations in physician contours of standard risk pelvic lymph nodes in patients receiving radiation therapy for prostate and nodal disease.
Methods: The contoured planning images from 152 patients treated with radiation therapy (RT) to the prostate and elective pelvic nodes from 2021-2024 were acquired. Patients were divided into 5 groups based on the original contouring physician and further divided into 8:2 (training:test) split.
Several fully convolutional neural network architectures were investigated for the task of voxel-wise segmentation of the elective pelvic nodes. Training perturbations investigated included uniform random noise distributions, left/right flipping, and random cropping.
Results: The AttentionNet architecture resulted in the best performance. The institutional model had median (standard deviation) DSC of 0.80 (0.07), median surface distance of 0.13mm (0.06mm), and max surface distance metrics of 3.3mm (6.7mm). Individual physician models had similar results. These are very similar to STAPLE CTV results of physician contouring in this region with DSC of 0.823 (0.07).
Conclusion: The delineation of nodal volumes presents a complex and intricate challenge, as individual physicians frequently exhibit systematic differences from their peers, which may or may not be clinically significant. These consistent discrepancies complicate the development of a singular 'ground truth' model for the segmentation task. We propose that it is more advantageous to create physician-specific models, which can more effectively highlight areas of divergence of a physician’s manual segmentation from their peers, an essential component of the peer review process.