# Efficacy of Automatic 3D Segmentation of the Upper Airway in CBCT or CT Scans via Artificial Intelligence Versus Manual Segmentation by Human Experts: A Systematic Review and Meta‐Analysis

## TL;DR

This study reviews and analyzes how well artificial intelligence performs compared to human experts in 3D segmentation of the upper airway from dental and medical scans.

## Contribution

The first meta-analysis evaluating AI's efficacy in upper airway segmentation compared to manual methods.

## Key findings

- AI showed precision, dice similarity score, intersection over union, and recall above 90%.
- Total volume difference was small but significantly above zero.
- Results suggest AI is promising but more studies are needed for conclusive evidence.

## Abstract

3D segmentation of the upper airway is crucial for dental and medical practices. However, it is a difficult and daunting task. Like almost all other areas, AI can theoretically help in airway segmentation. Nevertheless, AI's efficacy remains unknown. This meta‐analysis investigated this matter for the first time.

‎Various search engines/databases/articles were searched for articles published until April 25, 2025. All English‐language articles on the use of AI for upper airway evaluations based on CBCT or CT scans were included in the study. The desired population was considered studies assessing the upper airway. Intervention was the use of any tool of AI such as deep learning and machine learning for image analysis. The comparator was the manual analysis of CBCT or CT scans by human. The outcome was the analysis of upper airway on CBCT or CT images. The recorded and analyzed effect sizes were: accuracies, precisions, dice similarity scores, total volume differences, intersection over union (IoU), recall, or any other parameters relevant to segmentation. A meta‐analysis was conducted for each of the mentioned parameters if adequate data were available. The outcome was the analysis of upper airway on CBCT or CT images (PROSPERO: CRD42024508004).

Eleven studies were included, with 6 studies included in meta‐analyses. Most studies had a low risk of bias in most aspects. The qualitative part of review showed promising results for AI segmentation. Four of the effects sizes were meta‐analyzed: Precision,‎ dice similarity score, intersection over union, ‎ and recall were all above 90%.‎ Total volume difference was small but significantly above zero. Sensitivity analyses showed robustness of all meta‐analysis results. Publication bias was insignificant.

The results showed promising AI efficacies in 3D segmentation of the upper airway in CBCTs. However, much more studies are needed before decisive conclusions.