Artificial-Intelligence Diagnostic Accuracy for Radiographic Caries Detection: A Systematic Review

Objectives: Artificial Intelligence (AI), and machine learning (ML) a branch of AI, offers interesting promise for radiological caries detection. We aimed to answer: What is the diagnostic accuracy of AI/ML techniques for the radiographic detection of dental caries?
Methods: We performed a systematic review of diagnostic test accuracy(DTA) according to PRISMA-DTA recommendations. We included caries diagnostic-focused AI studies from 1999 to 2019 in PubMed, Scielo, EMBASE and arXiv using terms related to 'artificial-intelligence', 'machine-learning' or 'deep-learning' and 'caries'. References of included studies were also searched. We excluded studies that investigated image segmentation or any other objective rather than caries classification/detection. We selected studies that have sufficient information to construct contingency tables or that report these diagnostic measures. Risk of bias was assessed using QUADAS-2. Prospero registration:CRD42019125491.
Results: We identified 89 studies, of which 5 were included. Four studied provided enough data to estimate diagnostic accuracy, with reported sensitivity=81(74.5-86.1) and specificity=83(76.5-88.1). Due to high heterogeneity, we do not perform a meta-analysis. All studies had a high risk of bias and all were done on permanent teeth. In one study, Convolutional-Neural-Network(CNN) outperformed human-experts meanwhile in other dentists beat a Fully-Convolutional-Neural-Networks (FCNN)(accuracy of .78 vs .61) Another study compared different deep-learning techniques, with K-nearest Neighbors(k-NN), Large-Margin-Nearest-Neighbor(LM-NN) and Adaptive-Neural-Network(ADA-NN) better than Support-Vector-Machine(SVM) and Naive-Bayes(NB). Training dataset ranged from 40-2400 images.
Conclusions: The evidence of diagnostic precision of IA for the detection of caries is scarce. The most commonly used ML techniques was deep-learning with CNN. The reported diagnostic precision is medium-to-high (60-88%). In one study the IA surpassed human experts. There are several ML techniques used and the clinical utility of AI for caries detection is not yet well defined. There is no evidence of Therapeutic-, Patient-outcome- and Societal-efficacy.