Photo-Based Color Analysis in Restorative Dentistry: The Role of Artificial Intelligence Algorithms

This study aimed to develop and evaluate deep learning-based models for estimating CIELAB tooth color coordinates from standardized extraoral photographs, using in vivo spectrophotometric measurements as the reference standard, to provide practical and validating support for clinical shade selection in restorative dentistry.

A total of 1031 labeled and cropped tooth images were obtained from high-resolution extraoral photographs of 102 participants. In vivo CIELAB color coordinates (L*, a*, b*, C*, h°) were recorded from the maxillary anterior teeth using the VITA Easyshade V spectrophotometer. Five convolutional neural network (CNN) architectures (CustomCNN, ResNet18, EfficientNetB0, DenseNet121, and MobileNetV2) were trained using two types of input: preprocessed tooth photographs (cropped, resized, and normalized) and their corresponding anatomical positions (one-hot encoded). The performance of this combined-input model was evaluated using mean absolute error (MAE), coefficient of determination (R ²), and estimation accuracy based on the ΔE ≤ 2 clinical threshold. Intermodel performance differences were analyzed using one-way ANOVA (p < 0.001). The best-performing model was further trained to classify VITA Classical shade codes from spectrophotometric color coordinates to assess clinical applicability.

Among the models tested, MobileNetV2 achieved the best performance, with a MAE of 0.0593 and accuracy exceeding 94% when estimated CIELAB color coordinates from extraoral photographs were compared with in vivo spectrophotometric measurements (reference standard) using the ΔE ≤ 2 clinical threshold. No significant differences were found between MobileNetV2, DenseNet121, and ResNet18, and all three models significantly outperformed EfficientNetB0 and CustomCNN (p < 0.001). Estimation accuracy was highest for C*(chroma) and h°(hue) values, while slightly lower performance was observed for L*(lightness) and b* values. Anterior teeth exhibited higher accuracy than posterior teeth, likely due to more favorable image quality and light exposure.

AI-based models, particularly MobileNetV2, demonstrated high agreement with spectrophotometric reference measurements in estimating CIELAB tooth color coordinates from standardized extraoral photographs. These findings suggest their potential to support consistent and practical shade selection in esthetic restorative dentistry.

These results suggest that AI models have the potential to be applied in clinical shade selection procedures, providing a supportive approach that may contribute to improved consistency and assist decision-making in esthetic dentistry.