Detecção de dentes e classificação de patologias dentárias em imagens de radiografias panorâmicas utilizando aprendizado profundo

• Ano de defesa: 2025
• Autor(a) principal: MENDES, Alison Corrêa
• Orientador(a): QUINTANILHA, Darlan Bruno Pontes
• Banca de defesa: QUINTANILHA, Darlan Bruno Pontes , PAIVA, Anselmo Cardoso de , ALMEIDA, João Dallyson Sousa de , ALMEIDA, Luciana Salles Branco de , CRUZ, Luana Batista da
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Maranhão
• Programa de Pós-Graduação: PROGRAMA DE PÓS-GRADUAÇÃO EM CIÊNCIA DA COMPUTAÇÃO/CCET
• Departamento: DEPARTAMENTO DE INFORMÁTICA/CCET
• País: Brasil
• Palavras-chave em Português: radiografia panorâmica; aprendizado profundo; detecção de dentes; classificação de anomalias dentárias; redes neurais convolucionais;
• Palavras-chave em Inglês: panoramic radiography; deep learning; tooth detection; classification of dental anomalies; convolutional neural networks
• Área do conhecimento CNPq: Ciência da Computação
• Link de acesso: https://tedebc.ufma.br/jspui/handle/tede/6219

Panoramic radiography plays a fundamental role in dental diagnosis and treatment planning. However, its interpretation is subject to errors due to factors such as fatigue, stress, and the experience of the professional. This work proposes a deep learning-based method for the automatic detection of teeth and the classification of dental anomalies in panoramic radiographic images. Different convolutional neural network architectures were evaluated for both tasks. In tooth detection, the YOLOv8x architecture achieved the best results, with a mAP50 of 0.945 and a mAP50-95 of 0.578. In pathology classification, DenseNet121 achieved an F1 Score of 0.854, demonstrating a good balance between correct identification and the reduction of false positives. The results suggest that the proposed method, which separates the process into two stages using YOLOv8x for tooth detection and DenseNet121 for pathology classification, outperforms the approach where YOLOv8x alone is used to perform both tasks. While the proposed method achieved an F1 Score of 0.6014, the single-stage approach based solely on YOLOv8x obtained an F1 Score of 0.433. The method was trained and evaluated using the DENTEX dataset, which consists of panoramic radiographs annotated by specialists, including information on tooth numbering and the presence of pathologies such as impacted teeth, caries, and periapical lesions. These results reinforce the potential of the proposed approach as a decision-support tool in dental diagnostics, contributing to error reduction and improving the efficiency of panoramic radiograph analysis.