Ensemble learning methods with single and multi-model deep learning approaches for cephalometric landmark annotation

The study explores end-to-end deep learning frameworks and ensemble methods to enhance the accuracy of anatomical landmark identification in cephalometric radiographs, crucial for precise cephalometric analysis and effective orthodontic treatment planning. The methodology is strategically designed to address the complexities and variations in cephalometric images through a dual learning process: building base learners using end-to-end regression approaches and then constructing a meta-learner that ensembles the output of the base models to obtain collective predictions. Two primary approaches are followed to design base learners: an end-to-end single model and a multi-model strategy for localizing landmarks. Additionally, the efficacy of these approaches is examined by implementing two distinct deep architectures, namely UNet and UNetPlusPlus. The combination of the single model and multi-model approach with UNet and UNetPlusPlus results in four base learners, each displaying varying accuracy for different landmarks. To leverage the strengths of these individual models, three ensemble methods are explored, that includes supervised ensembling with XGBoost, unsupervised ensembling with the DBSCAN algorithm, and mathematical aggregation using weighted averaging. The methodology is evaluated on both public and private datasets, comprising 400 and 700 images, respectively. The multi-model approach demonstrates superior performance, achieving the highest detection accuracy of 79.9% (UNet) for the benchmark dataset and 92.8% (UNetPlusPlus) for the private dataset. The ensemble meta-model further boosts accuracy to 83.61% and 95.4%, respectively, reducing mean radial errors by 0.38 mm and 0.33 mm. These results highlight significant improvements in accuracy and error reduction through strategic combinations of deep learning architectures and ensemble techniques demonstrating the ability to significantly enhance cephalometric landmark annotation accuracy, which is critical for the practical applicability of the methodology.