Performance comparison of three artificial intelligence models in predicting gingival-colored porcelain compositions.

Statement of problem: Studies on the esthetic outcomes of soft tissue restoration coloration are lacking. Moreover, the relationship between ceramic powder proportions and their resulting color in gingival-colored restorations lacks investigation.

Purpose: The purpose of this in vitro study was to develop and compare 3 artificial intelligence-based systems-Residual Neural Network (ResNet), Multilayer Perceptron (MLP), and Genetic Algorithm-optimized Backpropagation (GA+BP)-for predicting gingival-colored porcelain compositions to improve color matching accuracy in restorative dentistry.

Material and methods: A total of 359 specimens were fabricated, including 286 standard and 73 extreme-proportion formulations. CIELab coordinates (L*, a*, b*) were measured using a dental spectrophotometer, and a database was established to correlate each gingival-colored porcelain powder composition with its corresponding CIELab values. Three models (ResNet, MLP, and GA+BP) were developed and evaluated using 5-fold cross-validation, with mean squared error (MSE) as the loss function. Performance metrics, including MSE, mean absolute error (MAE), explained variance, and training time, were statistically analyzed using the Kruskal-Wallis test followed by post hoc Dunn tests with Holm-Bonferroni correction. External validation used 10 new formulations, with ΔE₀₀ compared with perceptibility (<1.1) and acceptability (<2.8) thresholds via t tests or Wilcoxon tests (α=.05).

Results: ResNet achieved the lowest MSE of 0.0199 ±0.0003, outperforming MLP (0.0211 ±0.0003, P<.01) and GA+BP (0.0213 ±0.0002, P<.001). The model also demonstrated the lowest MAE (0.1069 ±0.0009), significantly lower than GA+BP (0.1086 ±0.0007, P=.002), but not MLP (0.1073 ±0.0004, P=.524). ResNet exhibited the highest explained variance (0.718 ±0.004), surpassing MLP (0.647 ±0.007, P<.05) and GA+BP (0.638 ±0.004, P<.001). GA+BP required the shortest training time (4.80 ±0.25 seconds per fold), less than MLP (5.62 ±0.30 seconds, P<.05) and ResNet (16.74 ±1.89 seconds, P<.001). In external validation, ResNet achieved an average ΔE₀₀ of 1.55 (95% CI: 1.14-1.95), lower than MLP (2.37; 95% CI: 1.72-3.02) and GA+BP (2.13; 95% CI: 1.54-2.72), with no significant difference among models (P>.05).

Conclusions: ResNet demonstrated the best accuracy in predicting gingival-colored porcelain compositions, as evidenced by the performance metrics. These findings support the use of artificial intelligence (AI)-driven systems, particularly ResNet, to enhance the accuracy and reproducibility of gingival color matching.