View Classification and Tooth Detection/Classification with Deep Learning in Intraoral Images

In this study, we utilize intraoral images for deep learning to provide valuable supports to dentists. This study unfolds in two main phases. Initially, ResNet152 is used to classify which part of the dental arch is captured in the intraoral image. Following that, we utilize YOLOv5s for teeth detection and tooth type classification. Additionally, this study discusses post-processing techniques applied after YOLOv5s. We fine-tuned ResNet152, originally pre-trained on the ImageNet dataset, with our dataset of 290 cases. Each case included a set of five views: a frontal view, two lateral views (left and right), and two occlusal views (upper and lower). Subsequently, we fine-tuned YOLOv5s, which was pre-trained on the Microsoft Common Objects in Context (MS COCO) dataset, with 215 cases for the tooth detection/classification. In the first phase, classifying the views into five categories with ResNet152 achieved a 100% accuracy rate. In the subsequent phase, complemented by the post-processing, the YOLOv5s models for tooth detection/classification yielded accuracy rates of 97% for the upper occlusal views, 94% for the lower occlusal views, and 87% for the frontal views. Our proposal outperformed the existing research in accuracy, promising significant clinical advancements.