Author: Rex A. Cardan, Carlos E. Cardenas, Quan Chen, Jingwei Duan, Joseph Harms, Joel A. Pogue, Richard A. Popple, Yi Rong, Dennis N. Stanley, Natalie N. Viscariello, Libing Zhu ๐จโ๐ฌ
Affiliation: Washington University in St. Louis, The University of Alabama at Birmingham, Mayo Clinic Arizona, University of Alabama at Birmingham ๐
Purpose: Manual verification of organs-at-risk(OARs) delineations is a critical yet time-intensive process, often susceptible to unintentional oversights. To assist the reviewing process, a population-based system that automatically detects and visualizes anomalous contour regions without requiring a patient-specific reference contour was developed.
Methods: A representative template contour, chosen for its suitability as an average segmentation, is used as common coordinate system for comparing all subsequent contours. Given a dataset, Bidirectional-Local-Distance(BLD) is employed to quantify the local distance disagreements between the test and template contours, following center-of-mass alignment. Population-based disagreements database for each point is then establish by analyzing BLDs for all points on the common reference. A new contour is flagged as outlier if 3% or more of points on the template contour fall outside the disagreement range[ยตโ3ฯ,ยต+3ฯ]. The majority voting strategy is utilized to generate the final prediction using three reviewed representative template contours. The disagreements database was established for 16 OARs from a retrospective dataset(n=3311), and tested on clinically-approved prospective dataset(n=1705) across CNS, Pelvis, and Abdomen sites.
Results: The method presented in this study allows the identification of delineation outliers while highlighting local anomalous regions. On the clinically-approved prospective dataset, 4.2%(n=58) of contours were flagged as outliers. The visualization capability facilitated rapid identification of these anomalous regions, with average time<15s. Manual review confirmed 81.0%(n=47) of the flagged outliers as true anomalies. Upon manual review, the identified outliers were primarily attributed to:(1) segmentation outliers, including over-contouring, under-contouring, and stray voxels; and(2) uncommon anatomical outliers, such as organ enucleation or tumor invasion.
Conclusion: The proposed system offers an automatic OAR segmentation outlier detection and visualization tool, without the need of a patient-specific reference. It has the potential to streamline the contour verification process. Ongoing investigations are underway to evaluate the clinical utility of this tool across a broader range of OARs.