A deep learning framework for automatic cephalometric landmark localization through gaussian heatmap regression

Abstract

Accurate cephalometric landmark localization is essential for orthodontic diagnosis, craniofacial assessment, and treatment planning. Yet reliable deep-learning performance on limited annotated datasets depends not only on network architecture but also on landmark representation, coordinate decoding, physical error computation, and validation design. This thesis develops a reproducible framework for the automatic localization of 19 anatomical landmarks in lateral cephalometric radiographs. The framework integrates a lightweight two-level U-Net with Gaussian heatmap regression, differentiable Soft-Argmax coordinate extraction, and image-specific anisotropic pixel-to-millimeter conversion. Evaluation used the PKU Cephalometric Dataset of 102 radiographs. The Holdout 70/30 protocol was defined as the primary evaluation, with 72 radiographs for model development and 30 fully isolated radiographs for final testing; a K-Fold 70/30 heatmap ensemble on the same 30-image external cohort and a Nested Leave-One-Out Cross-Validation (Nested LOOCV) procedure, providing one strictly out-of-sample prediction per radiograph, served as complementary validation. Under the primary protocol, the framework achieved a Mean Radial Error (MRE) of 2.1663 mm, with Successful Detection Rates of 80.35%, 87.37%, 91.40%, and 95.79% at thresholds of 2.0, 2.5, 3.0, and 4.0 mm, respectively. The K-Fold ensemble and Nested LOOCV yielded MREs of 2.2423 mm and 2.2541 mm (95% CI [1.2232, 3.2850]; median landmark error 1.1148 mm). Paired statistical analyzes on the 30 common radiographs revealed no statistically significant differences between protocols, supporting robust performance. The main contribution is methodological rather than architectural: established components are integrated into a coherent pipeline combining spatial prediction, continuous coordinate estimation, physically consistent error measurement, strict external evaluation, and matched statistical analysis. These results establish validation design as an integral component of deep-learning-based cephalometric localization on small annotated medical-imaging datasets.