Limitations of off-the-shelf multimodal foundation models in secondary caries detection on interproximal radiographs

To evaluate whether off-the-shelf general-purpose multimodal foundation models can achieve clinically meaningful diagnostic performance in detecting secondary caries on interproximal radiographs without domain-specific training, using GPT-5o (OpenAI) and Gemini 2.5 Flash (Google Inc.) as representative publicly accessible models. A comparative exploratory study used 50 interproximal radiographs of posterior teeth obtained from a dental imaging center, yielding 208 teeth. Each tooth was evaluated twice by each model in independent sessions, totaling 416 observations per model. A specialist in restorative dentistry established the reference standard. Models were assessed in a zero-shot setting using a standardized prompt. Diagnostic performance was evaluated at the tooth level using accuracy, sensitivity, specificity, precision, F1-score, and Cohen's kappa, with 95% confidence intervals estimated via bootstrap resampling (1000 iterations). Both models demonstrated limited diagnostic performance. GPT-5o showed moderate sensitivity (64.7%) but low specificity (32.3%), resulting in an accuracy of 41.3% and negligible agreement with the reference standard (κ =  - 0.021). Gemini 2.5 Flash exhibited higher sensitivity (82.8%) but extremely low specificity (13.0%), with lower accuracy (32.5%) and similarly negligible agreement (κ =  - 0.026). Both models showed a strong tendency toward false-positive classifications. Tooth localization performance was poor, with low identification rates in both models. GPT-5o and Gemini 2.5 Flash, when used in their off-the-shelf configurations, demonstrated inadequate diagnostic performance for detecting secondary caries on interproximal radiographs. Their limited accuracy, low specificity, and poor localization ability do not support their direct use for secondary caries detection on interproximal radiographs without domain-specific training and rigorous independent validation.