Deep learning for dental implant brand classification: a systematic review of imaging modalities, methodological variability, and model performance

Accurate identification of dental implant brands from radiographic images is clinically important, particularly when previous treatment records are unavailable. Deep learning (DL) models have shown promising performance for automated implant classification, yet it remains unclear how reliably this performance translates across different imaging modalities, datasets, and validation strategies. To systematically evaluate the diagnostic performance and methodological characteristics of deep learning models for dental implant brand classification using panoramic, periapical, and mixed radiographic datasets. This systematic review was reported in accordance with the PRISMA 2020 guidelines. PubMed, Scopus, Web of Science, and IEEE Xplore databases were searched for studies published between 2020 and 2026. Studies employing CNN-based, YOLO-based, transformer-based, AutoML, or hybrid deep learning models for implant brand classification were included. Extracted data included imaging modality, dataset characteristics, model architecture, validation strategy, and diagnostic performance metrics. Risk of bias was assessed using the QUADAS-2 tool. Because the performance metrics reported across the included studies were inconsistent and not statistically poolable, a structured narrative synthesis was performed in accordance with the Synthesis Without Meta-analysis (SWiM) guidelines rather than a meta-analysis. Twenty studies met the inclusion criteria. Reported diagnostic accuracy ranged approximately from 75% to 99%. Studies using smaller datasets, limited implant diversity, and manually cropped implant regions generally reported higher performance, whereas large multicenter datasets demonstrated comparatively lower but potentially more clinically realistic results. Panoramic radiographs were the most commonly used imaging modality, although periapical and mixed datasets also demonstrated strong diagnostic capability. Earlier studies primarily used transfer learning–based CNN architectures, while more recent investigations incorporated attention mechanisms, hybrid architectures, synthetic augmentation, and automated object-detection workflows. However, true geographically independent external validation remained uncommon. Deep learning models may demonstrate potential for dental implant brand classification. Nevertheless, diagnostic performance is influenced by dataset composition, implant diversity, preprocessing workflow, and validation strategy rather than model architecture alone. Future studies should prioritize multicenter external validation, standardized reporting, explainable AI integration, and clinically applicable automated workflows.