Development and evaluation of a deep learning model for occlusion classification in intraoral photographs.

Background: Malocclusion affects oral health and aesthetics, traditionally classified using systems like Angle's, which depend on physical exams or casts. Digital dentistry has shifted towards intraoral photography for documentation and assessment, though interpretation requires clinical expertise. The application of artificial intelligence (AI), and specifically deep learning, in medical imaging has been successful but remains largely unexplored in occlusal classification from intraoral photos. This study introduces a deep learning model to automate the classification of occlusal types from intraoral photographs, aiming to improve efficiency and objectivity in orthodontic diagnosis and treatment planning. Objectives: Occlusal classification is a crucial prerequisite for designing orthodontic treatment plans. Therefore, this study aims to develop an evaluation tool utilising a deep learning approach to automatically identify occlusal types reflected in digital oral photographs. Methods: Using a large-scale dataset with high-quality annotations (comprising 5,000 orthodontic intraoral photographs at a 45° lateral view and 2,200 at a 90° lateral view from 6,100 patients), three deep-learning models were developed based on Swin Transformer for the identification of various occlusal classifications: Molar occlusal relationships (M1, M2, M3), canine occlusal relationships (C1, C2, C3), and anterior overbite relationships (normal overbite, deep overbite, edge-to-edge bite, open bite, anterior crossbite, single-tooth crossbite or segmental crossbite). Results: Our model achieved weighted average F1-scores of 0.90 and 0.87 for molar and canine occlusal relationships, respectively. Regarding anterior overbite relationships, the model attained a weighted average f1-score of 0.89, with subclass F1-score ranging from 0.86 for edge-to-edge bite to 0.94 for deep overbite. Conclusions: Our deep learning model has successfully achieved the primary objectives of identifying molar and canine occlusal relationships, as well as anterior overbite relationships, using intraoral digital photographs. The demonstrated performance of this model highlights its potential for clinical applications. Clinical significance: The application of deep learning models for occlusal classification depicted in digital intraoral photographs, which enables clinicians to extract key information rapidly, holds significant implications for patient management and treatment monitoring in orthodontic practices.