Enhanced tooth segmentation algorithm for panoramic radiographs

• Autor(a) principal: Carneiro, José Andery
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações da USP
• Texto Completo: https://www.teses.usp.br/teses/disponiveis/59/59143/tde-20022024-073306/
• topic: Aprendizado profundo; Computer vision systems; Deep learning; Diagnóstico bucal; Oral diagnosis; Panoramic radiography; Radiografias panorâmicas; Segmentação de dentes; Sistemas de visão computacional; Teeth segmentation

Oral health encompasses a broad range of conditions, including dental caries, periodontal disease, tooth loss, and oral cancer. Maintaining optimal oral health requires both prevention and treatment of these conditions. Timely detection is crucial to prevent their progression. While clinical inspections are effective in many cases, they face limitations in identifying hidden or hard-to-reach issues. Dental radiography plays a vital role in ensuring accurate diagnoses. To enhance the speed and precision of radiograph analysis, oral health professionals are increasingly embracing advancements in Computer Vision, particularly leveraging Deep Learning for image processing. These techniques have given rise to various diagnostic tools, ranging from identifying cavities to classifying root canal treatments. A common initial step for these tools involves the detection of teeth in radiographic images. To enhance this critical phase, we introduce a modular system for teeth instance segmentation. This system comprises two key components: (i) dentomaxilo region detection (including mandible, maxilla and teeth) and (ii) segmentation of individual teeth within the identified dentomaxilo area. We employed RetinaNet for dentomaxilo region detection and Cascade Mask R-CNN for tooth identification. We trained these models using a dataset annotated by experienced professionals, which includes 935 panoramic radiographs with bounding boxes delimiting the dentomaxilo area and, among them, an additional 605 with tooth polygons, totaling 14,582 annotated teeth. These tasks are interconnected, with the output of one phase feeding into the next. Our system achieved good results, with dentomaxilo region detection scoring 92.446 mAP and 0.982 F1-score, and tooth segmentation attaining 79.222 mAP and 0.989 F1-score, surpassing benchmarks set by comparable studies. Our modular tool allows for future expansions, with the potential to integrate diverse new functionalities, such as tooth numbering or caries identification. Beyond serving as a diagnostic aid, offering support to dentists as a secondary opinion, our system has the potential to expedite the generation of epidemiological reports for large population samples.