Reliability-Aware View-Adaptive Consensus for 3D Cephalometric Landmark Identification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Manual 3D cephalometric analysis from CBCT is labor-intensive and prone to errors.
• Existing deep learning methods, like volumetric CNNs, are computationally expensive.
• Multi-view consensus methods can be unreliable due to view-dependent uncertainties.

Purpose of the Study:

• To develop an efficient and accurate framework for 3D cephalometric landmark identification from CBCT projection views.
• To address the limitations of manual annotation and existing automated methods.
• To improve the reliability and reduce computational demands in landmark identification.

Main Methods:

• Proposed a reliability-aware view-adaptive consensus framework for 3D landmark identification.
• Utilized a shared-weight 2D network to predict per-view 2D heatmaps and reliability scores.
• Employed an end-to-end differentiable geometric consensus for adaptive view fusion.

Main Results:

• Achieved a mean radial error of 1.26 mm with a 95% CI [1.20, 1.33].
• Reached a high successful detection rate of 88.41% at 2 mm.
• Demonstrated low computational loads and a favorable accuracy-efficiency trade-off.

Conclusions:

• The proposed framework offers an efficient and accurate solution for 3D cephalometric landmark identification from CBCT.
• Reliability-aware adaptive consensus improves landmark detection robustness and reduces variability.
• The method presents a practical alternative to computationally intensive approaches, with potential for clinical application.