Deep learning for cephalometric landmark detection: systematic review and meta-analysis.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Biomedical Engineering

Background:

• Deep learning (DL) is increasingly utilized for automated landmark detection in medical imaging, particularly for cephalometric analysis.
• Cephalometric analysis relies on accurate landmark identification for diagnosis and treatment planning.

Purpose of the Study:

• To systematically review and meta-analyze the diagnostic accuracy of DL for cephalometric landmark detection on 2-D and 3-D radiographs.
• To assess the quality of evidence supporting DL in cephalometric landmark detection.

Main Methods:

• Systematic review and meta-analysis of diagnostic accuracy studies (2015-2020) identified in Medline, Embase, IEEE, and arXiv.
• Included 19 studies primarily using convolutional neural networks on 2-D lateral radiographs.
• Assessed study quality using QUADAS-2, with a focus on risk of bias and applicability.

Main Results:

• DL models achieved a mean landmark prediction error around a 2-mm threshold.
• The proportion of landmarks detected within the 2-mm threshold was 0.799 (95% CI: 0.770 to 0.824).
• Most studies exhibited a high risk of bias and applicability concerns.

Conclusions:

• Deep learning shows promising accuracy for automated cephalometric landmark detection.
• Current evidence is consistent but limited by high risk of bias; further studies on generalizability and clinical utility are required.
• While 2-D imaging is well-studied, data on 3-D imaging are emerging but sparse.