Predicting Final Restoration Color Using Neural Network Models: The Impact of Substrate Lightness Versus Ceramic Shade, Translucency and Thickness

Abstract

Objectives: This study aimed to assess the influence of background color, ceramic shade, translucency, and thickness on the color matching of lithium disilicate restorations and to use a neural network model to predict the optimal parameters for shade matching. Methods: A neural network model was applied to evaluate the effects of lithium disilicate ceramic shade (A1, A2, A3), translucency (HT, M, T, LT), and thickness (0.5, 1.0, 1.5 mm), as well as background color (black, white, enamel and dentin shades A1–D4) on color matching. Color measurements (L*, a*, b*) were obtained using a spectrophotometer (CM-700d, Konica Minolta) in SCI mode. Results: Background color had a significant influence on color matching (p < 0.001, R2 = 0.740). The model explained 71.45% of the variance in final color values, with a mean absolute error (MAE) of 0.8578 units in the CIELab space. SHAP analysis identified initial L* value (42.32%), ceramic thickness (19.51%), and translucency (10.36%) as key predictors. Component-wise, L* had an R2 of 0.7594, a* had the lowest R2 (0.5993), and b* performed best (R2 = 0.7848). Conclusion: Background color and ceramic thickness are the most critical factors in minimizing color discrepancies in ceramic restorations. The developed model demonstrates promising potential in predicting the color outcomes of ceramic restorations. Clinical Significance: Neural networks offer promise as predictive tools for clinicians, aiding in selecting materials and fabrication parameters to achieve desired esthetic outcomes.