Related Experiment Video
Updated: Sep 9, 2026
Quasistatic Mechanical Testing for Computer-Aided Design and Manufacturing Occlusal Veneers Cemented to Milled Dentin Analog Material
Published on: December 20, 2024
Deep learning-based final color prediction and target-driven material screening for lithium disilicate ceramic
Xing Chen1, Shenglan Wu1, Mengyun Xu2
1Clinical Research Center for Oral Tissue Deficiency Diseases of Fujian Province & Fujian Key Laboratory of Oral Diseases & Fujian Provincial Engineering Research Center of Oral Biomaterial & Stomatological Key Laboratory of Fujian College and University, School and Hospital of Stomatology, Fujian Medical University, Fuzhou, Fujian, China.
Objectives:
To develop and evaluate a deep learning-based framework for predicting the final color of lithium disilicate ceramic veneers and to establish a target-driven screening algorithm for identifying and ranking candidate veneer-try-in paste combinations for various abutment-target shade scenarios.
Methods:
Lithium disilicate veneer specimens were fabricated in 7 shades and 3 thicknesses. Try-in pastes from 2 commercial systems and 29 VITA Linearguide 3D-MASTER shade tabs were used to simulate cementation and abutment/target shades, respectively. Spectral reflectance data were captured and converted to CIELAB coordinates. A fully connected neural network (FCNN) incorporating batch normalization and ReLU activation was trained to predict the final CIELAB coordinates of layered veneer assemblies from 12 material-related colorimetric inputs. Prediction accuracy was assessed using CIEDE2000 color differences (ΔE00), referenced against the perceptibility threshold (PT = 0.8) and acceptability threshold (AT = 1.8). The trained model was applied to 841 abutment-target shade scenarios to screen 210 candidate veneer-try-in paste combinations per scenario. The screened candidate combinations were further evaluated using measured final-color data.
Results:
In the test set, the FCNN achieved a mean ΔE00 of 0.50 ± 0.27, with 100% of the predictions falling below the AT and 87.50% below the PT. Clinically acceptable candidate combinations were identified in 267 scenarios, yielding 514 ranked candidate combinations. When evaluated using the measured final-color data, the mean ΔE00 was 0.55 ± 0.40 between measured and predicted colors, and 1.22 ± 0.51 between measured colors and target shades; 87.16% of the screened candidate combinations remained below the AT (ΔE00 < 1.8).
Conclusions:
The proposed FCNN-based framework accurately predicted veneer final color and enabled target-driven screening and ranking of candidate veneer-try-in paste combinations within the predefined material library.
Clinical Significance:
This framework may provide a clinically relevant decision-support approach by narrowing and prioritizing candidate veneer-try-in paste combinations before chairside try-in, potentially reducing reliance on empirical trial and error during material selection.