Multicenter Evaluation of a Deep Learning Model for Automatic Contouring of Cardiac Substructures in Adults.

PURPOSE: Efficient delineation of cardiac substructures is essential for assessing radiation-induced cardiac toxicity in thoracic radiotherapy. We clinically evaluated a deep learning model designed for automated segmentation of these structures. METHODS AND MATERIALS: The model built in collaboration with XXX (XXX) generated 21 cardiac substructures (whole heart, chambers, coronary arteries, valves, and great vessels). Five radiation oncologists rated raw contours on thirty contrast-enhanced CTs from two institutions using a 4-point Likert scale (3 = no correction, 2 = minor/moderate correction, 1 = major correction, 0 = not performed) and applied corrections when needed. Dice Similarity Coefficients (DSC) and 95th-percentile Hausdorff distances (HD95) were computed to compare model versus reference contours, clinician-corrected versus reference contours, and model versus STAPLE contours. RESULTS: Of 630 structures, 93.8% were scored either 3 or ≥2, and 6.2% received 1 by at least one clinician. For the coronary arteries (210 structures), 88.5% were scored either 3 or ≥ 2. Contour generation took approximately 1 min, with a median review and correction time of 7.0 min per CT (IQR, 5.8-9.8). The median DSC between the model and reference contours was 1.0 across all structures, while it ranged from 0.92 to 1.0 for clinician-corrected versus reference contours. Overall, HD95 was below 5 mm, for both clinicians versus reference and model versus STAPLE. CONCLUSIONS: The deep learning model provided satisfactory segmentation of cardiac substructures on contrast-enhanced CT, including coronary arteries. Its integration into a widely adopted treatment planning system lays the foundation for broad clinical implementation.