Deep learning to support the detection and classification of teeth, dental caries and restorations: a review

• Autor(a) principal: Carneiro, José Andery
• Data de Publicação: 2024
• Outros Autores: Zancan, Breno Augusto Guerra, Tirapelli, Camila, Macedo, Alessandra Alaniz
• Tipo de documento: Artigo
• Idioma: por
• Título da fonte: Brazilian Journal of Health Review
• Texto Completo: https://ojs.brazilianjournals.com.br/ojs/index.php/BJHR/article/view/68218
• topic: oral diagnosis; artificial intelligence; computer vision systems; dental caries; dental restoration

Oral health serves as a crucial indicator of individuals' overall well-being and quality of life, making it a pertinent concern regularly addressed by healthcare professionals. Utilizing imaging exams is imperative for detecting and identifying oral diseases and conditions, and the application of Artificial Intelligence (AI) has garnered attention for its potential in this realm. We conducted a systematic literature review focusing on the utilization of Deep Learning techniques in dental radiographs for the detection, segmentation, and classification of teeth, caries, and restorations. Our review encompassed automated searches across prominent databases including the ACM Digital Library, IEEE Xplore Digital Library, PubMed, and Scopus, yielding 393 primary papers published between 2012 and 2023. Following stringent inclusion and exclusion criteria, we thoroughly examined 68 papers, assessing their consistency and adequacy of different aspects in terms of the databases used, techniques implemented, and outcomes reported. It was noted that 41.66% of the analyzed papers lacked clear information regarding approval of data usage from ethics committees. Additionally, despite the interdisciplinary nature of computational techniques in oral health, 38.23% of the surveyed studies were conducted by teams comprising solely professionals from one specific area. Moreover, 66.18% of the papers focused solely on panoramic radiographs, with commonly utilized metrics including accuracy, recall, and precision. Notably, the U-Net and Mask R-CNN networks emerged as the most frequently applied methodologies. Despite the proliferation of investigations in this field, several challenges persist, including the limited availability of public datasets, inadequate detailing of developed methodologies, and a lack of systematization in result presentation. These challenges hinder a fair comparison between studies, presenting a significant obstacle to be addressed for further progress in the field.