Deep learning for caries detection: A systematic review

Objectives: Detecting caries lesions is challenging for dentists, and deep learning models may help practitioners to increase accuracy and reliability.We aimed to systematically review deep learning studies on caries detection.Data: We selected diagnostic accuracy studies that used deep learning models on dental imagery (including radiographs, photographs, optical coherence tomography images, near-infrared light transillumination images).The latest version of the quality assessment tool for diagnostic accuracy studies (QUADAS-2) tool was used for risk of bias assessment.Meta-analysis was not performed due to heterogeneity in the studies methods and their performance measurements.Sources: Databases (Medline via PubMed, Google Scholar, Scopus, Embase) and a repository (ArXiv) were screened for publications published after 2010, without any limitation on language.Study selection: From 252 potentially eligible references, 48 studies were assessed full-text and 42 included, using classification (n=26), object detection (n=6), or segmentation models (n=10).A wide range of performance metrics was used; image, object or pixel accuracy ranged between 68%-99%.The minority of studies (n=11) showed a low risk of biases in all domains, and 13 studies (31.0%) low risk for concerns regarding applicability.The accuracy of caries classification models varied, i.e. 71% to 96% on intra-oral photographs, 82% to 99.2% on periapical radiographs, 87.6% to 95.4% on bitewing radiographs, 68.0% to 78.0% on near-infrared transillumination images, 88.7% to 95.2% on optical coherence tomography images, and 86.1% to 96.1% on panoramic radiographs.Pooled diagnostic odds ratios varied from 2.27 to 32767.For detection and segmentation models, heterogeneity in reporting did not allow useful pooling.Conclusion: An increasing number of studies investigated caries detection using deep learning, with a diverse types of architectures being employed.Reported accuracy seems promising, while study and reporting quality are currently low.Clinical significance: Deep learning models can be considered as an assistant for decisions regarding the presence or absence of carious lesions.