Purpose: Manual contouring (MC) of patient organs remains a large time sink for physicians in radiation oncology. Artificial intelligence-based autocontouring systems (AI-AC) will likely reduce the time required by physicians when creating final contours. However, the increased reliance on AI-AC necessitates quantification of contour variance between physicians and the AI-AC in question. The work presented assesses accuracy of a commercially available AI-AC in comparison to physician-generated MC.
Methods: Contours were compared for 10 patients undergoing radiation treatment of the brain using MC and AI-AC. Contours for whole brain, brainstem, optic chiasm, and both optic nerves were generated using MC and the commercially available AI-Rad Companion Organs RT. Of note, physician-created “Whole Brain” structure was compared to the combination of the “Brain,” “Brainstem,” and “Optic Chiasm” substructures within AI-Rad Companion Organs RT. All structures were compared via Dice similarity coefficient (DSC) and mean distance to agreement (DTA).
Results: DSC ranged from 0.36 ± 0.22 to 0.98 ± 0.00 for the structures analyzed. DTA ranged from 0.76 ± 0.13 mm to 3.25 ± 0.99 mm for the structures analyzed. AI-AC of the head of a patient, including several structures not analyzed in this study, required approximately 2 minutes. Manually contouring the structures analyzed in this study is estimated to require approximately 15 minutes.
Conclusion: Issues were identified with AI-generated contours near the skull base, especially near the cribriform plate, and with gaps being created between the optic nerves and chiasm. These issues are of significant treatment concern and necessitate significant user review after AI-generated contours are complete. The largest variations were observed for the optic chiasm (DSC = 0.36) and Brainstem (DTA = 3.25mm). The current generation of AI-generated contours appear to create a useful head start for physician’s contours, but require significant editing based on anatomical interpretation inaccuracies and physician preference.
Segmentation, Anatomical Models, Brain