Federated Learning-Based CNN Models for Orthodontic Skeletal Classification and Diagnosis

Area of Science:

• Artificial Intelligence in Dentistry
• Medical Image Analysis
• Machine Learning for Orthodontics

Background:

• Accurate skeletal classification is crucial for orthodontic diagnosis and treatment planning.
• Current methods may lack efficiency or compromise patient data privacy.
• Federated learning (FL) offers a potential solution for collaborative model training without centralizing sensitive data.

Purpose of the Study:

• To evaluate the effectiveness of federated convolutional neural network (CNN) models for orthodontic skeletal classification.
• To compare the performance of FL against centralized learning (CL) and local learning (LL) frameworks.
• To assess the ability of FL to maintain data privacy while enabling collaborative model training.

Main Methods:

• Utilized DenseNet121 CNN architecture, enhanced with attention mechanisms and pooling blocks.
• Trained and evaluated models on cephalometric images from the ISBI and Dicle datasets.
• Benchmarked model performance using accuracy, sensitivity, and specificity across CL, LL, and FL frameworks.

Main Results:

• Federated CNN models achieved accuracy improvements exceeding 26% compared to baseline models.
• Augmented DenseNet121 models demonstrated significant performance gains under FL, comparable to CL.
• Specific models showed notable improvements, e.g., 20.86% gain over LL on the ISBI dataset.

Conclusions:

• Federated CNN models show significant potential for orthodontic skeletal classification.
• FL enables enhanced collaborative model training while preserving data privacy.
• This approach advances orthodontic diagnostics through secure, collaborative AI across institutions.