M. Farias
Texas State University,
United States
Keywords: AI, healthcare, dental images, diagnosis system
Summary:
We developed an opportunistic auxiliary diagnostic system designed to enable the early detection of systemic diseases using imaging exams routinely performed for other purposes. Dental imaging are simple, inexpensive exams, frequently conducted during standard dental visits. They represent an often-overlooked opportunity for the early medical referral of undiagnosed patients. Our approach focuses on panoramic radiographs, a two-dimensional dental examination that captures the entire oral cavity in a single image, encompassing not only the teeth but also the upper and lower jaws, surrounding tissues, and portions of the cervical spine. These exams can be leveraged for the screening and potential diagnosis of conditions such as cardiovascular disease and osteoporosis during routine dental check-ups. Artificial intelligence has already been applied to the diagnosis of these diseases; however, existing research still lacks methods capable of simultaneous multi-disease diagnosis and the use of advanced deep learning architectures that enhance the representation of medical image features. In this work, we designed a unified AI-based hybrid diagnostic framework that uses panoramic radiographs to opportunistically detect early signs of cardiovascular disease (e.g., atherosclerosis) and osteoporosis. The panoramic radiographs were systematically annotated by dental and radiology specialists for the presence of vascular calcifications and osteoporotic indicators, with ground truth established through bone mineral density (BMD) assessments and coronary angiography evaluations. For osteoporosis, the framework employed recent convolutional neural network (CNN) architectures trained to classify varying levels of disease severity. Their performance was evaluated against DXA scan results, the clinical gold standard for osteoporosis diagnosis. The findings demonstrated that the AI models outperformed general dentists on several metrics and closely approached the diagnostic performance of expert radiologists, underscoring the significant potential of AI-driven analysis of dental images for automated and accessible systemic disease screening in clinical practice. For cardiovascular disease, we implemented a hybrid AI architecture combining FasTVit, AttentionNet, and DC-UNet for the classification, detection, and segmentation of carotid artery calcifications (CACs). Two scenarios were evaluated: one in which detection and segmentation followed classification, and another in which the classification step was omitted. The classification model achieved accuracy, precision, recall, and specificity above 0.8, with the inclusion of the classification stage resulting in substantial performance improvements.