Integration of grey relational analysis and deep neural networks for optimized dental shade selection in aesthetic restorations

PURPOSE: To develop and validate an integrated system combining grey relational analysis (GRA) with deep neural networks (DNNs) for personalized oral aesthetic color optimization, and to evaluate its performance compared to traditional color selection methods in dental restoration. MATERIALS AND METHODS: Using a priori power analysis, 150 patients (70 males, 80 females; mean age 31.5 ± 7.2 years) were recruited for this study. For VITA shade classification, GRA was performed on 15 clinical variables (tooth color, facial/gingival parameters, demographics, preferences) which were weighted and these therefore formed the features of a DNN (15→128→64→32→16 neurons in 4 hidden layers) with 16-class output. The system exhibited a clinical accuracy of 93.2% against a consensus of prosthodontists (82.7%) on spectrophotometry alone. Technical accuracy (ΔE00 ≤ 2.0) was recorded as 94.7%. RESULTS: An integrated system was found to have an accuracy of 93.2% (95%CI: 90.6-95.8) for tooth color recommendation. This output was significantly better than the accuracy of visual assessment (68.5%) and spectrophotometric measurement (82.7%) (p < 0.001). Age-specific analysis revealed younger patients (20-35 years) preferred brighter shades (B1, A1), while older patients showed greater acceptance of natural color variations. This approach offers clinicians an evidence-based decision-making tool that improves both efficiency (processing time reduced from 180s-45s) and patient satisfaction (increased from 72.3%-91.4%). CONCLUSION: The combined application of GRA and DNN provides a recommendation support tool for optimizing the oral cavity's aesthetic color and designing personalized aesthetic treatment plans. Future studies in a multicenter setting would test generalizability.