Mandible-focused osteoporosis risk assessment using dental panoramic radiography and artificial intelligence models

• Ano de defesa: 2023
• Autor(a) principal: Machado, Leonardo Ferreira
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Artificial intelligence models; Avaliação de risco; Dental panoramic radiography; Modelos de inteligência artificial; Osteoporose; Osteoporosis; Radiografia panorâmica dentária; Risk assessment
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/59/59135/tde-17082023-112055/

Osteoporosis is a systemic disease that provokes bone mineral density (BMD) losses that eventually cause severe bone fractures. Since it is a silent disease, many are diagnosed only after fractures. Several opportunistic image-based diagnoses are being investigated in combination with artificial intelligence (AI) models in the literature. The present study proposes a mandible-focused dental panoramic X-ray image (PAN) analyses using artificial intelligence models to assess the osteoporosis disease risk. To accomplish that, we initially developed an automatic mandible segmentation for PAN images using an ensemble of deep learning algorithms. To develop this mandible segmentation model, we used two datasets: an in-house dataset (IHD) prepared with 393 PAN images manually annotated by a specialist and a third-party dataset composed of 116 images previously annotated. U-Net and HRNet architectures were considered individually and an ensemble format with and without segmentation post processing. With this approach we achieved the best mandible segmentation performance in the literature with 98.2%, 97.6%, 97.2%, accuracy, dice similarity, and intersection over union, respectively. In the second moment of this study, we used this algorithm to extract the mandible image region of interest (ROI) from PAN images from 380 PAN images from patients who also underwent bone mineral density (BMD) examination. Those patients were organized into two groups according to WHO criteria for diagnosis: healthy (no signs of osteoporosis) and disease risk (osteopenia and osteoporosis). We trained the EfficientNetV2-L model using I) the entire PAN images as inputs and II) the mandible segmentation ROI to separate these two groups. We observed that the model using the mandible segmentation achieved better accuracy and recall (73.9% and 83.0%) than the models trained with the entire image, which indicates considerable gains of using this mandible-focused approach.