Performance of a residual neural network system versus dental technicians for gingival porcelain shade matching.

PURPOSE: This study aimed to compare the accuracy of a residual neural network (ResNet) system with experienced dental technicians for gingival shade prediction. MATERIALS AND METHODS: CIELab (L*, a*, b*) coordinates were measured from three adjacent 1 × 3 mm2 zones within the gingival region of 18 volunteers using a spectrophotometer (Crystaleye, Olympus). Zirconia-based prosthetic specimens (1.5-mm thickness) for the upper right central incisor were digitally designed and milled, with one fabricated using conventional visual shade matching (CZR; Kuraray Noritake Dental) by experienced technicians and the other using a ResNet-based shade prediction system. ΔE00 values were calculated relative to natural gingiva using the CIEDE2000 formula. ΔE00 were statistically analyzed using paired t-tests (overall comparison) and repeated measures analyses (regional comparison). One-sample t-tests compared values to acceptability (AT = 2.8) and perceptibility (PT = 1.1) thresholds (α = 0.05). RESULTS: The ResNet-based system demonstrated significantly lower overall ΔE00 (4.169 ± 2.048) compared to technicians (5.625 ± 1.967; p < 0.001), although both exceeded the acceptability threshold (AT = 2.8). Regionally, ResNet outperformed technicians in the middle (3.486 ± 1.310 vs. 5.724 ± 2.074; p < 0.001) and lower zones (3.509 ± 2.142 vs. 5.023 ± 1.883; p = 0.007), but not in the upper zone (5.511 ± 1.978 vs. 6.129 ± 1.886; p = 0.317). CONCLUSIONS: The ResNet-based system demonstrated statistically better shade-matching performance compared to experienced dental technicians for gingival shade matching, both in subjective evaluation and objective ΔE00 measurement, particularly in the middle and lower gingival zones. However, ΔE00 values for both methods exceeded clinical acceptability.