Deep learning for imaging diagnosis of jaw cystic lesions and maxillofacial tumors: A narrative review

Abstract

Deep learning continues to advance imaging-based diagnosis in oral and maxillofacial radiology. This narrative review has synthesized recent deep learning applications for detecting, classifying, and segmenting jaw cystic lesions and maxillofacial tumors on panoramic radiographs and cone-beam computed tomography scans. It has summarized representative one-stage detectors and convolutional neural network/transformer-based classifiers, along with segmentation methods, reported performance metrics, and key use-case considerations. In addition to this synthesis, the review has critically examined dataset constraints, spectrum and site bias, device-related heterogeneity, annotation inconsistency, and gaps in model explainability as well as described how these limitations restrict generalizability. Practical considerations for clinical implementation are also discussed, including workflow placement, quality assurance, and governance, followed by emerging research directions such as federated learning, multimodal fusion, and radiomics–deep learning combinations, each evaluated in terms of feasibility and current evidence maturity. Key evaluation metrics are interpreted in the context of dental imaging. Overall, current findings suggest that deep learning may enhance early and consistent recognition of jaw lesions, support surgical planning through automated delineation, and promote standardized interpretation, provided that models undergo external validation, reporting remains transparent, and deployment is guided by appropriate clinical oversight.