Deep learning-based localization of bounded edentulous spaces in intraoral occlusal images

ObjectivesThis study developed a deep learning pipeline to localize bounded edentulous spaces with missing teeth FDI number directly from intraoral photographs, eliminating reliance on radiographs or controlled imaging. Materials and MethodsA total of 4,373 intraoral occlusal images were included in the study. Of these, 367 images with at least one bounded edentulous space were held out for pipeline evaluation. The remaining 4,006 images were split into training (80%), validation (10%), and test (10%) sets for model development. The pipeline included ResNet-18, YOLOv8m and ResNet-101. Bounded edentulous were inferred based on FDI discontinuities and localized via spatial interpolation. ResultsThe pipeline was evaluated on a test set of 367 occlusal images, comprising 4,964 expected teeth. This equated to a precision of 88.2%, recall of 89.1%, an F1-score of 88.6%, and MCC of 0.875. The pipeline showed stable diagnostic performance for actual use cases and made full inference in 9 seconds per image on consumer-grade hardware, indicating its viability for use with standard clinical workflows. ConclusionsThe pipeline showed reliable ability to detect and localize bounded edentulous spaces. Its consistent performance across a wide range of different clinical cases justifies its use as a diagnostic tool in routine dental practice. Clinical SignificanceOur results show that accurate detection of edentulous spaces is now possible even from unstandardized clinical images and is less dependent on radiographs and manual charting. This simplifies workflows in day-to-day practice, aids precision in orthodontic and implant planning, and makes AI tools more accessible for use in those practices with minimal image infrastructure.