Deep Learning-based Assessment of the Relation Between the Third Molar and Mandibular Canal on Panoramic Radiographs using Local, Centralized, and Federated Learning

TL;DR

This study compares centralized, federated, and local deep learning approaches for classifying the relationship between the third molar and mandibular canal in panoramic radiographs, highlighting the trade-offs between performance and data privacy.

Contribution

It demonstrates that federated learning outperforms local learning and approaches centralized training performance in dental radiograph classification tasks.

Findings

Centralized learning achieved the highest AUC of 0.831.

Federated learning showed intermediate performance with AUC 0.757.

Local learning generalized poorly across clients with AUC range 0.619-0.734.

Abstract

Impaction of the mandibular third molar in proximity to the mandibular canal increases the risk of inferior alveolar nerve injury. Panoramic radiography is routinely used to assess this relationship. Automated classification of molar-canal overlap could support clinical triage and reduce unnecessary CBCT referrals, while federated learning (FL) enables multi-center collaboration without sharing patient data. We compared Local Learning (LL), FL, and Centralized Learning (CL) for binary overlap/no-overlap classification on cropped panoramic radiographs partitioned across eight independent labelers. A pretrained ResNet-34 was trained under each paradigm and evaluated using per-client metrics with locally optimized thresholds and pooled test performance with a global threshold. Performance was assessed using area under the receiver operating characteristic curve (AUC) and threshold-based metrics, alongside training dynamics, Grad-CAM visualizations, and server-side aggregate monitoring signals. On the test set, CL achieved the highest performance (AUC 0.831; accuracy = 0.782), FL showed intermediate performance (AUC 0.757; accuracy = 0.703), and LL generalized poorly across clients (AUC range = 0.619-0.734; mean = 0.672). Training curves suggested overfitting, particularly in LL models, and Grad-CAM indicated more anatomically focused attention in CL and FL. Overall, centralized training provided the strongest performance, while FL offers a privacy-preserving alternative that outperforms LL.