Multiclass Semantic Segmentation of Dental Radiographs for Dental Ailments Using Deep Learning

Abstract

Recognition of Dental disease using Artificial intelligence has become an eminent focus of research. Different dental radiographic imaging procedures have empowered dental professionals to diagnose various dental conditions. This paper investigates root canal treated tooth, dental caries, and impacted tooth by segmenting dental panoramic radiographs. The root canal treated tooth is segmented to assess the prognosis of endodontic treated tooth, to rule out any periapical pathology like cyst, periapical granuloma, to detect bone resorptions around root canal treated tooth and in furcation areas due to missed out lateral canals, to evaluate the fit of gutta percha (under extension, over extension in relation to radiographic root apex). Dental caries is a prevalent infectious disease affecting the human race and it is considered to have a significant effect on health and personality. The radiographic examination is extremely valuable in identifying carious lesions that are difficult to identify through meticulous, in-depth clinical examination. The radiographic assessment of tooth caries is useful to diagnose and restore the affected tooth. The impacted teeth is segmented to check the eruption level of the third maxillary and mandibular molars. This work explores the CNN architecture U net by substituting the encoder block with backbone architectures such as Resnet34, Efficientnetb0, InceptionV3, and VGG16 through transfer learning. The performance of Resnet34 is impressive than other neural networks. The performance metric IoU and f1-score was evaluated for each model.