Health insurance plans provide access to medical and dental treatments at predictable costs, democratizing healthcare services that would otherwise be financially inaccessible to a significant portion of the population. To ensure the sustainability of this model, dental audits validate performed procedures by analyzing supporting images submitted by service providers. The identification of inconsistencies in these images, such as reuse, manipulation, or duplication, may result in claim denials and other contractual penalties. This study aims to evaluate computational methodologies for automatic detection of similar dental images to assist the audit process in identifying potential image duplication fraud. The methodology employed consisted of a systematic comparison of various visual feature extraction models, including pre-trained convolutional neural network architectures (VGG-16, ResNet-50, MobileNetV2, DenseNet-121, ShuffleNetV2) and the Bag of Visual Words (BoVW) method, combined with different similarity metrics (Euclidean distance, Chebyshev distance, and cosine similarity). The study was developed in two stages: initially, the models were evaluated on public dental datasets; subsequently, experiments were conducted with real data from Hapvida NotreDame Intermédica (HNDI), incorporating specific audit rules for qualifying suspicions. The experimental results demonstrated that the combination of the BoVW model with cosine similarity and audit rules presented superior performance, achieving a Top-1 precision of 95.65% in identifying duplicate images. The implementation of these techniques is expected to provide a significant increase in dental audit productivity and coverage, contributing to greater efficiency in resource allocation and economic sustainability of supplementary health systems.