AI Automates Airway Segmentation on CBCT in Sleep-Disordered Breathing — High Accuracy, Limited Clinical Validation

AI (mostly U-Net CNNs) segments craniofacial airways on CBCT with Dice scores 0.90–0.97 and ICC >0.90 vs. manual methods across 14 studies. But external validation is scarce and only 1 study tested OSA prediction against polysomnography — not yet ready for routine use.

• Systematic review (PRISMA 2020/PROSPERO) of 14 studies; deep learning architectures — chiefly U-Net and SpatialConfiguration-Net — dominated; datasets, targets, and metrics varied considerably (high heterogeneity).
• AI substantially reduced annotation time vs. manual segmentation (exact time savings not reported in source); volumetric ICC consistently >0.90 indicates strong agreement.
• Critical limitation: evidence is almost entirely internal technical validation; external validation was limited and only one study evaluated clinical OSA prediction against polysomnography as a reference standard.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.