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

Cone-beam computed tomography (CBCT)-based three-dimensional (3D) cephalometric analysis relies on accurate anatomical landmark identification, yet manual annotation is time-consuming and subject to inter- and intra-observer variability. While volumetric convolutional neural networks can improve accuracy, their computational and memory demands limit practical deployment. Multi-view consensus offers an efficient alternative by predicting per-view two-dimensional (2D) heatmaps and fusing them geometrically, but its performance can degrade when view-dependent uncertainty produces unreliable results. We propose a reliability-aware view-adaptive consensus framework for 3D cephalometric landmark identification from CBCT projection views. A shared-weight 2D network predicts per-view 2D heatmaps and landmark reliability scores, which adaptively modulate each view's contribution in an end-to-end differentiable geometric consensus. This framework yields deterministic fusion without stochastic inlier sampling or multi-stage refinement. With 5-fold cross-validation, the proposed method achieved the lowest mean radial error (1.26 mm; 95 $\%$ CI [1.20, 1.33]) and the highest successful detection rate at 2 mm (88.41 $\%$ ), while maintaining low computational loads. Ablation studies further validated the key design choices and highlighted a favorable accuracy-efficiency trade-off with a limited number of views.