Acurácia diagnóstica de um sistema de inteligência artificial na detecção de lesões de cárie proximal: estudo retrospectivo com correlação clínico-radiográfica

Abstract

Dental caries remains the most prevalent oral health condition worldwide, with the early detection of proximal lesions representing a clinical challenge due to the limitations of visual examination and the inherent subjectivity of radiographic interpretation. In this context, artificial intelligence has been proposed as an auxiliary tool to enhance diagnostic precision and reduce inter-examiner variability. This study aimed to evaluate the diagnostic accuracy of an artificial intelligence system for detecting proximal caries lesions on digital bitewing radiographs, comparing its performance with that of human examiners and verifying the correspondence between radiographic lesion depth and therapeutic decisions recorded in dental charts. This was a retrospective, cross-sectional, analytical diagnostic accuracy study. The sample comprised 1,245 proximal surfaces of posterior teeth from 284 digital bitewing radiographs of 182 patients treated at the Universidade Federal do Ceará, Campus Sobral, obtained between January and December 2025. Images were analyzed by the Overjet® system, based on deep learning algorithms and convolutional neural networks, using the radiographic criteria of the American Dental Association (ADA) harmonized with the ICDAS (International Caries Detection and Assessment System) for lesion classification (E0–D3). The reference standard was established by consensus between two Oral Radiology specialists. Agreement between the artificial intelligence and the radiologists' consensus was moderate (κ = 0.598), a value numerically close to that observed between the radiologists themselves in independent assessment (κ = 0.580), suggesting that the system's diagnostic variability remained within the range expected among experienced examiners. The Restorative Dentistry specialists demonstrated near-perfect agreement among themselves (κ = 0.972). For global proximal caries detection, the artificial intelligence yielded sensitivity of 72.4%, specificity of 97.1%, accuracy of 83.9%, and an area under the ROC curve of 0.847 (95% CI: 0.829–0.866). For lesions with dentinal involvement, accuracy increased to 92.4% and the area under the ROC curve to 0.871. Analysis of the correspondence between radiographic depth and therapeutic decisions revealed a statistically significant therapeutic gradient [χ²(10) = 474; p < 0.001; contingency coefficient = 0.525], with a progressive increase in the frequency of invasive treatment from 0.9% in sound surfaces to 68.2% in lesions reaching the inner third of dentin. The proportion of surfaces with no recorded clinical decision was 61.1% of the total evaluated. It is concluded that the artificial intelligence system demonstrated diagnostic performance compatible with the variability observed among human specialists, with high specificity and good discriminative ability, particularly for lesions with dentinal involvement. The significant association between radiographic depth and therapeutic decisions reinforces the clinical relevance of the tool as an instrument to support evidence-based decision-making aligned with the principles of minimally invasive dentistry.