Can artificial intelligence estimate L*, a*, b* values of teeth from intraoral photographs?
TL;DR
Both AI models deviated significantly from spectrophotometric measurements and did not achieve clinically precise L*, a*, and b* values, and ChatGPT-5.1 Plus and Gemini 3 Pro showed model-dependent color interpretation patterns influenced by photographic conditions.
Abstract
STATEMENT OF PROBLEM Accurate shade selection is critical for natural dental esthetics, but, despite their objectivity, spectrophotometers (SPM) are limited in clinical use because of cost and practicality. Although multimodal artificial intelligence (AI) models may serve as accessible alternatives, their ability to estimate L*, a*, and b* values from standardized intraoral photographs remains uncertain.
Purpose
The purpose of this study was to evaluate the performance of ChatGPT-5.1 Plus and Gemini 3 Pro in comparing CIELab values obtained from intraoral photographs using different photographic systems compared with reference measurements obtained using an SPM. MATERIAL AND
Methods
Twenty participants (n=20) with a healthy maxillary right central incisor were included. Standardized intraoral photographs were obtained using a digital single-lens reflex (DSLR) camera with twin flash (TF), DSLR with ring flash (RF), and a smartphone (SP) with a mobile dental photography (MDP) light system under 5500 K cross-polarized illumination. All photographs were calibrated using a gray reference card. Each AI model generated 3 repeated outputs per photograph; the mean of these outputs was used for statistical analysis. Data were analyzed using ANOVA with Tukey post hoc tests and independent-samples t tests (α=.05).
Results
The CIELab values estimated by ChatGPT-5.1 Plus and Gemini 3 Pro from photographs obtained using different photographic systems differed significantly from the spectrophotometric reference measurements and from each other (P<.001). In both models, the highest L* values were observed in the SPM group and the lowest in the DSLR+RF group (P<.001). The a* and b* values varied according to the photographic system and AI model, with statistical similarity to the SPM observed in selected groups. Overall, ChatGPT-5.1 Plus and Gemini 3 Pro produced significantly different mean values across all color parameters, with P values ranging from <.001 to .026.
Conclusions
Both AI models deviated significantly from spectrophotometric measurements and did not achieve clinically precise L*, a*, and b* values. ChatGPT-5.1 Plus and Gemini 3 Pro showed model-dependent color interpretation patterns influenced by photographic conditions. Standardized photographic protocols and calibrated datasets are necessary to improve the clinical applicability of AI-based shade estimation.