Efficacy of artificial intelligence in detecting dental caries in interproximal radiographies: a systematic review with meta-analysis

uma revisão sistemática com metanálise

Abstract

Introduction: Deep learning consists of a form of artificial intelligence based on the neural structure of the brain that has shown excellent performance in carrying out variable tasks and dominant robotic language learning. In dentistry, deep learning has also stood out, especially in dental radiology, helping in the diagnosis process of the most diverse pathologies. In this process, tooth decay is one of the oldest and most important oral health problems. Despite this, even today, its diagnosis remains a challenge for the dentist. Currently, technical variables can assist the clinical examination in this process and the use of x-rays is one of the main and highly accurate complementary exams commonly used. A recent systematic review found that there was no difference in the detection of tooth decay using deep learning when compared to the assessment of a specialist in dental radiology. However, this review evaluated it through periapical radiographs. Taking into account that interproximal radiographs are better indicated when the carious process is smaller or in the interproximal regions, a systematic review is necessary to compare this technique based on this examination, which is the objective of this research. Methodology: The guiding clinical question of this systematic review was: Is convolutional deep learning effective in detecting dental caries through bitewing radiographs? The acronym PIRO was used in order to establish the eligibility criteria for comparative studies in which P = population (interproximal radiographs of deciduous or permanent teeth), I = Index test (Deep learning CNN), R = Reference standard (radiographic interpretation by specialist), O = Outcome (detection of dental caries), S = Study design (Comparative studies (observational or clinical trials). A search was carried out in the electronic databases Medline (Pubmed), Cochrane Central Registry of Controlled Trials ( CENTRAL), Virtual Health Library (VHL), EMBASE and Science Direct, using the descriptors (deep learning) AND (dental caries). Additionally, a search in the gray literature was checked using Google Scholar for the first 50 results found. The reference list of the included studies and literature or systematic reviews of similar topics were checked in order to find potential eligible articles.There was no restriction regarding the language or year of publication of the studies. The entire search, study selection and data collection process was carried out by two independent researchers. After the electronic search, the references were exported to the EndNote X8 reference organizing software (Clarivate Analytics, PA, USA). Duplicates were excluded and references were screened using titles and abstracts. Then, selected studies were searched for full reading and evaluated according to the eligibility criteria. The risk of bias of the included studies was assessed using the QUADAS-2 tool. The synthesis of results from the included primary studies was presented through meta-analysis using the Meta-disc software that compared the performance of each deep learning to the reference standard. Results: The electronic search resulted in 331 references, of which nine publications met the eligibility criteria and were included in the systematic review. The included articles contributed to a total sample of 5,930 bitewing radiographs. In assessing the risk of bias, most studies were judged as “low” for the patient selection domain. All studies obtained “uncertain” and “low” judgments for the index test and reference standard domains, respectively. The flow and time domain obtained “low” risk of bias in most studies. In assessing applicability, the studies were judged as “low” and “high” risk of bias in the patient selection domain. A “low” risk of bias judgment was obtained in all studies for the index test and reference standard domains. Six studies presented primary data that allowed the meta-analysis to be carried out. The observed results indicated sensitivity of 0.90 (CI: 0.88-0.91, I²: 99.3%), specificity of 0.25 (CI: 0.25-0.26, I-square: 99, 9%), negative likelihood ratio (LR-) of 0.21 (CI: 0.18-0.24, I²: 95.4%) and positive likelihood ratio (LR+) of 1.38 (CI: 1 .36-1.39, I²: 99.9%) . Conclusion: Deep learning appears to be a tool with excellent specificity and good sensitivity in detecting dental caries through bitewing radiographs. In this way, artificial intelligence through the deep learning technique proves to be an excellent ally for professionals in the diagnostic detection of tooth decay through interproximal radiographic examinations. Well-designed studies are encouraged in order to better corroborate the results achieved here.