Abstract
Background/Objectives: Accurate skeletal classification is essential for orthodontic diagnosis. This study evaluates the effectiveness of federated convolutional neural network (CNN) models for skeletal classification using cephalometric images from the ISBI and Dicle datasets. This research aims to evaluate the effectiveness of federated learning (FL) for orthodontic skeletal classification by comparing its performance against centralized learning (CL) and local learning (LL). The objective is to determine whether FL can achieve competitive performance while preserving data privacy and enabling collaborative model training across multiple institutions. Methods: The DenseNet121 model and its augmented versions, incorporating channel attention, spatial attention, squeeze and excitation, and spatial pyramid pooling blocks, are proposed and adapted for the study. Models are evaluated on the ISBI and Dicle datasets using accuracy, sensitivity, and specificity metrics, with performance gains benchmarked across CL, LL, and FL frameworks. Results: Accuracy improvements exceed 26% compared to the baseline model on FL framework. The DenseNet121_SA model, augmented with spatial pyramid pooling blocks, achieves a 20.86% performance gain over LL settings on the ISBI dataset. Similarly, the DenseNet121_SA model, augmented with spatial attention, and DenseNet121_SA_SE model, augmented with spatial attention and squeeze and excitation, obtain 16.58% and 15.22% by not sacrificing performance loss with respect to CL. The inclusion of the Dicle dataset provides additional validation for the models. Conclusions: Federated CNN models exhibit significant promise for orthodontic skeletal classification. These models demonstrate the potential of FL to enhance collaborative model training while preserving data privacy. This approach represents a step forward in leveraging precise orthodontic diagnostics technology by enabling a data-secure collaborative artificial intelligence among various orthodontic clinics.