EXACTITUD DIAGNÓSTICA DE MODELOS DE INTELIGENCIA ARTIFICIAL PARA DETECTAR CARIES INTERPROXIMALES EN RADIOGRAFÍAS BITEWING USANDO ICCMS COMO ESTÁNDAR DE REFERENCIA: REVISIÓN SISTEMÁTICA

Abstract

Objective: To evaluate the diagnostic accuracy of artificial intelligence models for detecting or classifying interproximal caries on bitewing radiographs using ICCMS radiographic criteria as the reference method for diagnostic labeling. Materials and methods: A systematic review of diagnostic accuracy studies was conducted following the PRISMADTA and PRISMA 2020 guidelines. Original studies evaluating AI models applied to bitewing radiographs for the detection or classification of interproximal caries using ICCMS criteria were included. The search was performed on March 2, 2026, in PubMed, Scopus, and ScienceDirect. Study selection, data extraction, and methodological assessment were carried out by two independent reviewers. Risk of bias and applicability were assessed using QUADAS-2. Due to methodological heterogeneity, a descriptive synthesis was performed. Results: Five studies published between 2022 and 2025 were included. Most evaluated YOLO-based architectures, and one study assessed pretrained software, with a total of 6,384 bitewing radiographs before data augmentation. The models showed variable diagnostic performance and reported clinical and computational metrics, including sensitivity, specificity, accuracy, AUC, precision, F1-score, recall, average precision (AP), and mean average precision (mAP). A meta-analysis could not be performed because of differences in study designs, models, units of analysis, and reported metrics. Conclusion: AI should be considered a support tool for radiographic interpretation, not an autonomous diagnostic method. The available evidence is limited by methodological heterogeneity, lack of external validation, and incomplete reporting of results.