A Hybrid Transfer Learning Approach to Teeth Diagnosis Using Orthopantomogram Radiographs

Abstract

The rise in the emphasis on oral diseases has elevated the need to automate the diagnostic process of such diseases. Fortunately, the availability of modern computing devices has made the automated diagnosis of teeth readily possible using deep learning. Despite this, concerns about the accuracy and function of automated diagnosis remain among patients. To showcase the performance of such algorithms, we propose two approaches for the task of teeth diagnosis utilizing Orthopantomograms (panoramic radiographs): 1) a direct classification approach; and 2) a hybrid approach that combines a deep learning model with a traditional classifier. The results revealed that all ten chosen deep learning models experienced a similar or improved performance when used in conjunction with a machine learning classifier. In particular, Vision Transformer (ViT) performed the best with a record accuracy of 96% using both the direct and hybrid approaches. However, the hybrid framework combining AlexNet with a Support Vector Machine achieved an accuracy of 94%, and although it falls short of ViT in terms of performance, it comprises far fewer parameters. This highlights the approach’s effectiveness in improving performance without the need to use a deeper model, making it well-suited for clinical adoption where efficiency is important.

Keywords

• Dental imaging
• dental informatics
• deep learning
• machine learning
• orthopantomography
• transfer learning