CeLDA+: Prototypical learning for age-robust cephalometric landmark detection

Accurate cephalometric landmark detection is essential for orthodontic diagnosis. However, existing methods predominantly focus on adults and overlook adolescents, whose developmental variations, such as unerupted teeth and mixed dentition, introduce substantial appearance changes and degrade detection performance. These differences motivate a unified framework that performs reliably for adolescent and adult cases. To address this challenge, we propose CeLDA+ (Cephalometric Landmark Detection across Ages), a prototypical learning framework for landmark detection in adolescent and adult populations. Rather than relying on highly variable local appearances, it formulates landmark detection as semantic matching in a high-dimensional feature space. By mapping morphologically diverse instances to compact regions around shared landmark prototypes, the model promotes consistent localization for adolescent and adult cases without requiring age-specific designs. We further introduce two modules: Prototype Geometry Regularization, which enforces geometric consistency among landmarks, and Prototype Relation Mining, which captures semantic dependencies between anatomically related structures. For comprehensive evaluation, we construct two large-scale, multi-center datasets annotated with 46 and 201 landmarks, comprising 2950 samples in total and representing the largest cephalometric benchmarks to date. CeLDA+ consistently outperforms prior state-of-the-art methods across all three benchmarks. On the two mixed-age benchmarks, it achieves the best detection accuracy on both adolescent and adult subsets under joint training, while maintaining the lowest computational cost. CeLDA+ also performs well in two downstream clinical evaluations: skeletal classification, following prior benchmark practice, and cephalometric tracing analysis, newly introduced in this work to provide contour-level evaluation for cephalometric landmark detection. These findings further highlight its applicability in real-world orthodontic scenarios. Our source code and benchmark datasets are available at GitHub.