Automated Differentiation of Caries Requiring Filling and Caries Necessitating Root Canal Treatment Using Machine Learning

Abstract

ObjectivesNovice dentists or those at the onset of their professional careers require assistance in diagnosing cases that necessitate fillings or root canal treatments. In this study, we propose an innovative recommendation system based on deep learning to assist dentists in identifying the type of tooth caries (requiring filling and necessitating root canal treatment) and determining the appropriate treatment for detected caries. Correctly identifying the type of caries in teeth with no caries, only one type of caries, or more than one type of caries is important for determining the type of treatment to be applied.MethodsWe utilized 1253 bitewing images augmented with various variations, employing three different segmentation methods to automatically detect caries types in the first molar teeth. Furthermore, this study introduces a novel recommendation system for determining the treatment type required for the detected caries type, which represents a significant contribution to this field. The YOLOv8, U-Net, and Detectron-2 networks were evaluated for their efficacy in detecting various types of caries and recommending appropriate treatment methods.ResultsThe pixel-label-based comparative results generated by these methods on data labeled by experienced dentists were as follows: 95.03% for Detectron2, 90.88% for U-Net, and 89.23% for YOLOv8. The determination of the type of caries and the recommendation of the type of treatment differ from each other. In terms of treatment recommendations, the success rates of the three methods were as follows: Detectron-2, 88.09%; YOLOv8 70.23%; and U-Net, 61.90%. Consequently, Detectron-2 produced the most successful outcome among the three methods. These results are acceptable for the auxiliary treatment recommendation system.ConclusionThe system can serve as a supportive tool for less-experienced dentists and as a diagnostic aid for experienced practitioners. The learning-based segmentation method shows great promise for clinical use in the recommendation of treatment.