A prospective multicenter trial of deep learning auto-segmentation for organs at risk in thoracic radiotherapy.

Abstract

Widespread clinical implementation of rapidly evolving auto-segmentation tools remains constrained by a scarcity of high-quality prospective evidence. Here we show the results of a prospective, multicenter, observational trial (NCT05787522) evaluating the clinical performance of a deep learning model (iCurveE) for artificial intelligence (AI)-assisted delineation of organs at risk (OARs) in thoracic and breast cancer radiotherapy. Computed tomography images from 500 patients across five centers are annotated by 37 physicians using manual, AI-generated, and AI-assisted methods. Eleven thoracic OARs are evaluated based on the primary endpoints of volumetric Dice similarity coefficient (vDSC) and contouring time, alongside secondary metrics including 95% Hausdorff Distance (HD95). We prospectively annotate 2,483 OAR sets (27,043 OARs): 993 manual, 497 AI-generated, and 993 AI-assisted. AI-assisted delineation achieves significantly better vDSC (mean, 0.902) and HD95 (mean, 5.20 mm) than manual delineation (mean vDSC, 0.857; mean HD95, 8.01 mm; p < 0.0001) while improving time efficiency by 81.63% (median: 10.0 vs. 55.0 min; p < 0.0001). AI-assisted delineation reduces performance variability across centers and physicians with varying expertise. This study validates the clinical applicability of AI-assisted delineation in improving delineation performance and promoting healthcare equity.