2988 - A Data-Driven Quality Assurance for Auto-Segmentation Model Updates Using a Localized Anisotropic Margin Metric

Purpose/Objective(s): Auto-segmentation models require periodic updates, yet commonly used overlap metrics (e.g., DSC) can miss clinically meaningful localized boundary errors and are confounded by organ size. We propose a novel approach using baseline data to establish anisotropic variations, enabling a robust, automated non-inferiority test for model updates.

Materials/Methods: The clinical baseline model was commissioned using 36 patients (male pelvis and head-and-neck) with 21 types of OAR. The performance of established baseline model was quantified by the anisotropic surface deviations from reference contours using deformable point cloud registration- BLD (DPCR-BLD). Anisotropic safety variation was defined at each point under a Gaussian assumption (µ±2s, 95% interval) of baseline model to reference deviations across the cohort. This established a point-specific Anisotropic Surface Dice Similarity Coefficient (Ani-SDSC) threshold reflecting commissioned standards. The deviation of updated model from the baseline model was subsequently evaluated against these point-wise anisotropic variations.