DEEP LEARNING-BASED DETECTION OF INTRABONY AND FURCATION DEFECTS ON PERIAPICAL RADIOGRAPHS: SYSTEMATIC REVIEW

Abstract

Background: Periodontal disease is a leading cause of tooth loss, with intrabony and furcation defects representing advanced stages that critically impact treatment planning and prognosis. Accurate radiographic detection remains challenging due to interpretive variability and the limitations of two-dimensional imaging. Deep learning offers a promising approach to automate and enhance the detection of these defects on widely used periapical radiographs.

Methods: A systematic review was conducted following a comprehensive search of electronic databases (PubMed, Scopus, Web of Science, IEEE Xplore, Google Scholar) for studies applying deep learning to detect intrabony and/or furcation defects on periapical radiographs. Eligibility criteria included original research using deep learning models. Data on study design, model architecture, dataset characteristics, and performance metrics were extracted and synthesized qualitatively.

Results: Eight studies met the inclusion criteria, with five primary experimental studies included in the synthesis. Studies demonstrated considerable heterogeneity in design and methodology. Performance metrics varied, with furcation defect detection consistently achieving higher accuracy and AUC values (e.g., up to 94.97% accuracy, AUC up to 0.868) compared to intrabony defect classification, which showed more moderate performance (e.g., mAP@0.5 of 0.504, AUC of 0.77). Convolutional neural networks were the predominant architecture. Key influencing factors included dataset size, annotation quality, defect morphology, and model design.

Conclusion: Deep learning models show significant potential, particularly for the detection of furcation involvement on periapical radiographs. However, performance for intrabony defects remains more variable and challenging. To advance towards clinical application, future research requires standardized methodologies, larger and more diverse datasets, robust external validation, and consideration of integration into clinical workflows.