Evaluation of an AI-Based Model for Proximal Caries Detection on Bitewing

The detection of proximal caries on bitewing radiographs is a common clinical problem that is often difficult to make an accurate diagnosis, and even among experienced clinicians, interobserver variation has been reported as high as 20–30%. The study aimed to assess the accuracy of a deep learning-based AI model in proximal caries diagnosis on bitewing radiographs and its performance compared with that of the general dental practitioner and oral radiologist. The sample comprised 1,240 bitewing radiographs with 4,960 proximal surfaces of patients treated from 2020-2023. All images were independently assessed by two calibrated oral radiologists and consensus interpretations were used as the reference standard. Carious lesions were classified based on depth of the lesion into enamel (E1–E2) or dentin (D1–D3). This AI model was based on a modified ResNet-50 convolutional neural network which was pre-trained on the ImageNet dataset and fine-tuned on 9800 annotated dental radiographs. Diagnostic performance was evaluated by calculating the sensitivity, specificity, positive predictive value (PPV) and the area under the receiver operating characteristic curve (AUC). Cohen's kappa was used to determine interrater agreement. The AI model has a sensitivity of 87.3%, specificity of 91.6%, PPV of 84.1% and AUC of 0.94. The mean sensitivity and specificity for the general dental practitioners were 71.4% and 88.9% (AUC: 0.81) respectively, and for the oral radiologists were 90.2% and 93.1% (AUC: 0.96) respectively. The AI model had significantly better accuracy than GP in identifying early enamel lesions (74.1% vs. 51.3%, p<0.01), and close to the specialist level of accuracy. There was substantial agreement between the AI model and the reference standard (Cohen's kappa, 0.79). The results indicate that AI-based radiographic analysis could offer a reliable support for proximal caries detection, especially in the early stages in general dental practice