[Accuracy of multi-task network based on vision Transformer in the three-dimensional upper airway analysis]

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Computational Anatomy

Context:

• Accurate three-dimensional (3D) upper airway analysis is crucial for diagnosing and treating various medical conditions.
• Traditional segmentation methods can be time-consuming and subjective.
• Developing automated, accurate segmentation tools is a key area of research.

Purpose:

• To evaluate the accuracy and clinical applicability of a multi-task vision Transformer model for segmenting the 3D upper airway and its subregions.
• To compare the model's performance against established software (3D Slicer and Dolphin 3D) as gold standards.
• To assess the model's consistency and agreement with gold standards using Bland-Altman analysis and intraclass correlation coefficient (ICC).

Summary:

• A multi-task vision Transformer model was developed for automatic segmentation and volume measurement of the 3D upper airway and pharyngeal subregions using cone-beam CT (CBCT) data.
• The model demonstrated high accuracy (ICC=0.97) for overall upper airway segmentation compared to 3D Slicer.
• Segmentation of the nasopharynx, velopharynx, and glossopharynx showed good agreement with Dolphin 3D (ICC=0.94-0.96), but hypopharynx segmentation was less accurate (ICC=0.69).

Impact:

• The findings suggest that the multi-task vision Transformer model has potential for clinical application in upper airway analysis.
• The model's performance indicates a need for further improvements in robustness and generalization, particularly for the hypopharynx.
• This research contributes to the advancement of AI-driven tools for quantitative analysis in orthodontics and related fields.