Artificial Intelligence Based Auto-Contouring for Organs at Risk in Head and Neck πŸ“

Author: Mylinh Dang, Laila A Gharzai, Xinlei Mi, Poonam Yadav πŸ‘¨β€πŸ”¬

Affiliation: Northwestern Feinberg School of Medicine, Northwestern Memorial Hospital, Northwestern University Feinberg School of Medicine, Northwestern Medicine 🌍

Abstract:

Purpose: Delineation of organs at risk (OAR) in the head/neck region requires substantial physician time. Many artificial intelligence (AI) based auto-contouring software are commercially available. The purpose of this study is to evaluate the currently available AI contouring tools for clinical use.
Methods: Twenty patients with head/neck cancer were retrospectively identified and images were anonymized for this analysis. Two commercially available software, Limbus Contour (version 1.6.0) and MIM, were utilized to generate software-defined AI OARs. Gold standard OARs were defined by a head/neck radiation oncologist at the time of the patients’ initial treatment: mandible, brachial plexus, brain, brainstem, oral cavity, cochlea, submandibular glands, lips, optic structures, parotids. Accuracy and fidelity (measured as Jaccard index (JI), Dice similarity coefficient (DSC), Hausdorff distance (HD), and mean distance to agreement (MDA)) were calculated between gold standard and AI contours.
Results: A total of 229 contours generated by Limbus and MIM were compared. Overall, Limbus performed significantly better than MIM in JI (0.69 vs 0.63), p<0.001), DSC (0.82 vs 0.77, p<0.001), HD (8 vs 10, p<0.001), MDA (1.09 vs 1.5, p<0.001) for all OARs. For individual OARs, Limbus contour was better than MIM for most OARs, including brainstem, oral cavity, cochlea, submandibular gland, optic structures, and parotids. MIM outperformed Limbus for some structures: mandible (JI p=0.005, Dice p=0.006), brain (JI p=0.002, Dice p=0.002, MDA p=0.002), lips (Dice p=0.03); for brachial plexus, no significant difference in accuracy/fidelity was noted between the two. Limbus also contours additional structures as compared to MIM including esophagus, larynx, pharyngeal constrictors, and pituitary gland.
Conclusion: When compared to physician-drawn gold standard OARs, the commercially available Limbus contour software performed significantly better than MIM in the head/neck region. Limbus contour also provides additional OAR contours as compared to MIM. Further work exploring commercially available software in other disease sites is needed.

Back to List