The Evaluation of a Deep Learning Approach to Automatic Segmentation of Teeth and Shade Guides for Tooth Shade Matching Using the SAM2 Algorithm

Accurate shade matching is essential in restorative and prosthetic dentistry yet remains difficult due to subjectivity in visual assessments. We develop and evaluate a deep learning approach for the simultaneous segmentation of natural teeth and shade guides in intraoral photographs using four fine-tuned variants of Segment Anything Model 2 (SAM2: tiny, small, base plus, and large) and a UNet baseline trained under the same protocol. The spatial performance was assessed using the Dice Similarity Coefficient (DSC), the Intersection over the Union (IoU), and the 95th-percentile Hausdorff distance normalized by the ground-truth equivalent diameter (HD95). The color consistency within masks was quantified by the coefficient of variation (CV) of the CIELAB components (L*, a*, b*). The perceptual color difference was measured using CIEDE2000 (ΔE00). On a held-out test set, all SAM2 variants achieved a high overlap accuracy; SAM2-large performed best (DSC: 0.987 ± 0.006; IoU: 0.975 ± 0.012; HD95: 1.25 ± 1.80%), followed by SAM2-small (0.987 ± 0.008; 0.974 ± 0.014; 2.96 ± 11.03%), SAM2-base plus (0.985 ± 0.011; 0.971 ± 0.021; 1.71 ± 3.28%), and SAM2-tiny (0.979 ± 0.015; 0.959 ± 0.028; 6.16 ± 11.17%). UNet reached a DSC = 0.972 ± 0.020, an IoU = 0.947 ± 0.035, and an HD95 = 6.54 ± 16.35%. The CV distributions for all of the prediction models closely matched the ground truth (e.g., GT L*: 0.164 ± 0.040; UNet: 0.144 ± 0.028; SAM2-small: 0.164 ± 0.038; SAM2-base plus: 0.162 ± 0.039). The full-mask ΔE00 was low across models, with the summary statistics reported as the median (mean ± SD): UNet: 0.325 (0.487 ± 0.364); SAM2-tiny: 0.162 (0.410 ± 0.665); SAM2-small: 0.078 (0.126 ± 0.166); SAM2-base plus: 0.072 (0.198 ± 0.417); SAM2-large: 0.065 (0.167 ± 0.257). These ΔE00 values lie well below the ≈1 just noticeable difference threshold on average, indicating close chromatic agreement between the predictions and annotations. Within a single dataset and training protocol, fine-tuned SAM2, especially its larger variants, provides robust spatial accuracy, boundary reliability, and color fidelity suitable for clinical shade-matching workflows, while UNet offers a competitive convolutional baseline. These results indicate technical feasibility rather than clinical validation; broader baselines and external, multi-center evaluations are needed to determine its suitability for routine shade-matching workflows.