Automated Identification of AI Contouring Outliers Using Template Shape-Based Analysis and Centroid Mapping without Case-Specific Ground Truth
Abstract
Purpose
Minor and major errors may occur in both manual and automatic organ contouring process, such as mis-localization or severe under/over-segmentation. They can subsequently impact the accuracy of dose optimization and planning. This study aims to develop an automated detection method for identifying major contouring errors using a combination of template-based shape analysis and centroid mapping method.
Methods
Prospective contours were aligned to template shapes using Coherent Point Drift (CPD) to identify shape outliers, while Random Sample Consensus (RANSAC)-based centroid mapping was used to detect location outliers. OAR shapes and centroid templates were derived from clinically approved cases. 4,339 AI-segmented cases and 1,825 clinically approved cases were included in this study. If the bilateral distance or centroid distance between the prospective and template contours exceeded a defined threshold, the case was flagged as a major error. The threshold was tuned using the raw AI-generated contours that contained various errors and was subsequently applied to clinically approved contours for evaluation.
Results
The template-based shape alignment approach efficiently detected shape outliers, with 19 cases in the heart, 41 in the brainstem, 12 in the left kidney, and 145 in the liver. In clinically approved contours, the number of flagged cases for location outliers was 196 cases with 2σ threshold. Based on expert review, the true positive (TP) cases reached stable values after threshold adjustment, at 20 cases for clinical approved contours.
Conclusion
The developed workflow proved effective in identifying significant contouring errors and providing automated alerts for clinicians prior to manual review. Our findings provide guidance for selecting an appropriate threshold based on the available time and resources for review. This approach has the potential to reduce clinical workload and improve efficiency and accuracy in radiation therapy contouring.