Convolutional neural network for automated identification of periodontal bone loss on panoramic radiographs: A task-based narrative review

Abstract

Objectives: This review aimed to evaluate the performance of convolutional neural networks (CNN) in the automated identification of periodontal bone loss (PBL) on panoramic radiographs. Review: This task-based narrative review was developed from a structured literature search of PubMed, Scopus, ScienceDirect, and reference lists for full-text studies published in 2013-2023 that applied CNN-based methods to PBL on panoramic radiographs. Of 336 records, 34 duplicates were removed, 302 records were screened, 14 full texts were assessed, and 11 core studies were included. Ten studies incorporated classification, nine detection, and five segmentation. Reported performance was generally promising, but direct comparison was limited by heterogeneous datasets, disease definitions, architectures, data splits, and outcome metrics. Classification accuracy in studies reporting this metric was approximately 76-95%; segmentation studies reported high Dice and intersection-over-union values in selected tasks. Newer studies published after the original search show continued movement toward multicenter datasets, periodontitis staging, defect-pattern analysis, and longitudinal assessment. Conclusion: CNN shows strong potential for the accurate and rapid automated identification of PBL on panoramic radiographs through classification, detection, and segmentation approaches, and may assist dentists in diagnosing periodontal disease; however, small dataset size, limited image quality, and the absence of accompanying clinical data remain limitations that should be addressed in future research.