Implementation and Clinical Evaluation of AI-Based Auto-Contouring for Organ at Risk Segmentation
Abstract
Purpose
To assess the clinical implementation, contour accuracy, and workflow efficiency of a commercially available deep learning auto-contouring system for CT-based radiotherapy planning across multiple anatomical sites.
Methods
AI-Rad Companion Organs RT was implemented as a cloud-hosted solution across two clinical sites. Automated DICOM routing enabled direct import of AI-generated organ-at-risk contours into the treatment planning system. Thirty CT datasets representing head and neck, lung, abdomen, prostate, and breast cases were retrospectively evaluated. Manual contours created by a certified dosimetrist were used as the reference standard. Workflow efficiency was quantified by comparing total contouring time, including segmentation and physician review, between manual and AI-assisted workflows. Contour quality was evaluated using physician qualitative scoring and quantitative metrics, including Dice similarity coefficient and mean Hausdorff
Results
AI-assisted contouring reduced total contouring time in nearly all cases. Mean time savings were 79% (range: 52-83%) for abdomen, 66% (range: 59–72%) for lung, 58% (range: 45–67%) for head and neck, and 27% (range: 15–41%) for prostate cases. Breast cases demonstrated smaller and more variable efficiency gains, with a mean reduction of 17% (range: -40–42%), influenced by a single outlier. Overall, 89.7% of AI-generated contours were rated as clinically acceptable with minor or no edits. Across 310 structures, the mean Dice similarity coefficient was 0.82 and the mean Hausdorff distance was 2.10 mm. Higher geometric agreement was observed for larger organs, with lower Dice values noted in anatomically complex regions.
Conclusion
AI-Rad Companion Organs RT was successfully integrated into routine clinical workflows and demonstrated substantial reductions in contouring time while maintaining clinically acceptable contour accuracy across multiple anatomical sites. These findings support the clinical utility of AI auto-contouring as a tool to improve segmentation efficiency when used with appropriate clinician oversight.