Deep learning for osteoporosis screening in dental practice: a systematic review

To conduct a systematic review evaluating the performance of deep learning (DL) tools in osteoporosis screening using dental imaging and to assess whether these models have been implemented in dental practice.

Search was performed across 7 electronic databases and 2 sources of grey literature. Included studies applied DL algorithms to dental radiographs or CT in adults diagnosed with osteoporosis, using dual-energy X-ray absorptiometry (DXA) or expert assessment as the reference standard. Risk of bias was assessed using the QUADAS-2 (Quality Assessment of Diagnostic Accuracy Studies 2) tool, alongside an AI-specific checklist and GRADE to evaluate certainty of evidence.

Thirteen studies met the inclusion criteria. DXA was the most common reference standard (n = 08), and panoramic radiography was the predominant imaging modality (n = 12). Accuracy was the most frequently reported metric (n = 12). Most models used pretrained convolutional neural networks, such as VGG16, GoogleNet, ResNet, and AlexNet. ResNet and EfficientNet architectures showed superior performance, particularly when combined with ensemble techniques and clinical covariates. However, no study reported external validation or implementation in dental practice, limiting their applicability.

All DL models demonstrated potential as supportive tools for osteoporosis screening in dentistry. However, the absence of external validation and clinical integration limits their real-world use. Future research should focus on standardization and development of accessible, validated systems.

This systematic review is the first to show that, despite advancements in DL for osteoporosis screening, clinical applicability remains limited. It underscores the need for robust, user-friendly interfaces to facilitate integration into routine dental practice.