ADVANCEMENTS IN ARTIFICIAL INTELLIGENCE AND COMPUTER VISION FOR DENTAL IMAGING ANALYSIS: SELF-SUPERVISED LEARNING INNOVATIONS

Abstract

This dissertation explores the application of self-supervised learning methods in dental radiology to address the challenges posed by limited data availability for training deep learning models. The overarching goal is to enhance the efficiency and accuracy of automated systems for various dental diagnostic tasks, including teeth numbering, detection of dental restorations, orthodontic appliances, implant systems, marginal bone level, and dental caries from panoramic radiographs, CBCT images, intra-oral 3D scans, and dental radiographs. Key contributions include the development of several novel approaches: Self-supervised Learning for Dental Panoramic Radiographs: Utilizing SimMIM and UM-MAE with Swin Transformer, we achieved significant improvements in teeth detection and instance segmentation, increasing the average precision by 13.4% and 12.8%, respectively, over baseline methods. Self-Distillation Enhanced Self-supervised Learning (SD-SimMIM): Enhancing SimMIM with self-distillation loss, we improved performance on teeth numbering, dental restoration detection, and orthodontic appliance detection tasks, demonstrating superior outcomes compared to other methods. DentalMAE for Intra-oral 3D Scans: Extending the mesh masked autoencoder (MeshMAE), DentalMAE evaluates predicted deep embeddings of masked mesh triangles, yielding better generalization and higher accuracy in teeth segmentation tasks. DEMAE for Dental CBCT Images: Proposing the Deep Embedding MAE (DEMAE), which measures the closeness of predicted deep embeddings of masked patches to their originals, we achieved significant accuracy improvements in teeth segmentation from CBCT images. Masked Deep Embedding (MDE) for Implant Detection: By leveraging MIM, we developed MDE to enhance dental implant detection, creating a comprehensive Implant Design Dataset (IDD) with expert annotations, significantly boosting detection performance. Deep Embedding of Patches (DEP) for Bone Loss Assessment: An extension of MAE, DEP improved the accuracy of marginal bone level detection, supported by the creation of a Bone Loss Assessment Dataset (BLAD) with detailed annotations. Masked Deep Embedding of Patches (MDEP) for Caries Detection: This method enhanced dental caries detection performance, validated on the CariesXrays dataset, demonstrating higher precision and recall rates compared to traditional baselines. Through these innovations, the dissertation establishes the efficacy of self-supervised learning in overcoming data scarcity in dental imaging, offering promising AI-driven solutions for improved diagnostics and patient care in dentistry.