Evaluation of Image Derived CIELAB Colour Metrics for Machine Learning Classification of ICDAS Caries Severity: A Multimodal In Vitro Study

Objective: The aim of this study was to determine whether smartphone- and intraoral camera-derived CIELAB parameters reproduce spectrophotometric colour variation across ICDAS defined occlusal caries severity, agree with spectrophotometric measurements, and support machine learning classification and cross-modality transfer. Methods: A total of 298 extracted posterior teeth were classified as sound (ICDAS 0), initial (ICDAS 1–2), or moderate (ICDAS 3–4) lesions. Spectrophotometric measurements were obtained from five predefined occlusal sites per tooth. Each tooth was also imaged using a smartphone and intraoral camera under standardised conditions, and five corresponding regions of interest were analysed for image-derived CIELAB values. CIELAB patterns and agreement were assessed using Linear Mixed-Effects models, Intraclass Correlation Coefficients (ICCs), and Bland–Altman analysis. Nineteen tooth-level colour and spatial variation features were evaluated using five classifiers with stratified fivefold tooth-level cross-validation and identical fold assignments across modalities. Cross-modality transfer was assessed using spectrophotometer-trained models applied to aligned image-derived features. Results: Spectrophotometry showed the most consistent severity-related decrease in L* and increases in a* and b*. Smartphone measurements broadly reproduced the L* pattern and differentiated severity through a*, whereas intraoral cameras showed limited separation for L* and b*. Agreement was device- and coordinate-dependent; the highest ICCs were observed for smartphone L* (0.785) and intraoral camera a* (0.844), while systematic bias and wide limits of agreement indicated non-interchangeability. Smartphone-derived features achieved the highest balanced accuracy (0.836–0.918), followed by the intraoral camera (0.639–0.664) and spectrophotometry (0.443–0.552). Cross-modality transfer remained partial, with prediction agreement of 54.0–74.5% (k = 0.263–0.541). Conclusions: Image-derived CIELAB features captured lesion-related colour information but were not directly interchangeable with spectrophotometry. Image derived features performed best when the model was trained and tested using data from the same measurement method. Cross-device generalisability across different camera systems remains to be established.