A Multi-Stage Framework for Automated Cephalometric Landmark Detection: Addressing Resolution and Spatial Invariance

Abstract

Cephalometric analysis is a cornerstone of modern orthodontics, enabling precise diagnosis and treatment planning by identifying key morphometric landmarks. Since manual identification of these landmarks is labor-intensive and prone to variability, developing automated systems has been essential. Existing automated solutions often fall short of clinical requirements due to challenges such as small, insufficiently diverse datasets, poor generalizability, and computational inefficiency. Addressing these gaps, we propose three distinct methodologies for automated cephalometric landmark detection, each addressing specific limitations and building upon the strengths of the previous one. The frameworks leverage the Aariz dataset, known for its diverse resolutions and spatial variations. The final, most optimized methodology proposes a three-stage cascaded framework that seamlessly integrates global and local feature learning while ensuring invariance to resolution and spatial dimensions. In the first stage, the craniofacial region is extracted by predicting the locations of edge landmarks, achieving spatial invariance. This identified region is then fed to the second stage, where a multi-scale CNN performs simultaneous landmark prediction through heatmap regression, introducing resolution invariance and enhancing global feature learning. These pseudo-predictions are then used to generate high-resolution crops from the original image around the predicted landmark locations. The final stage refines the coarse landmark predictions for all landmarks using a single model optimized for local feature learning, further enhancing computational efficiency. The framework demonstrates robust generalizability across diverse cephalograms, achieving clinically acceptable performance on multiple datasets. On the Aariz dataset, our method achieved an overall mean radial error (MRE) of 1.69 ± 3.36 mm with 81.18% landmarks falling within the clinically acceptable range of 2 mm. These results highlight the framework’s reliability for landmark detection and its success in handling diverse cephalograms, positioning it as a practical solution for clinical use that streamlines workflows and enables timely, accurate treatment planning.