IADR Abstract Archives
Digital Shade-Matching Using Back-Propagation Neural Network Model
Objectives: Color assessment and reproduction have been the most challenging tasks in esthetic dentistry. To satisfy patients’ esthetic demands, dentists and dental technicians dedicate themselves to provide natural-looking restorations. However, due to the complicated optical characteristics of natural teeth, harmonious color matching of artificial restorations with natural dentition is still difficult to achieve. The present study was aimed at integrating the back-propagation neural network (BPNN) into the process of digital shade matching.
Methods: In this study, CIE L*a*b* values were retrieved from the shade guide (Vita 3D). To train the BPNN, each color shade tab was measured six times to form a dataset of 90 values. The training process randomly selected 63 values from the dataset as the training data and the remaining values as the testing data. We constructed the BPNN with Python and used the Keras framework. The BPNN architecture consisted of three input variables, two hidden layers, and four output variables. The parameters of the BPNN, including the number of hidden layer nodes, activation function, optimizer, and batch size, were determined by a systematic test. A tooth color was arbitrarily picked and passed through the trained BPNN to obtain the predicted output. Finally, the color differences(ΔE) between the real tooth color and the prediction were calculated.
Results: The optimal parameters of the BPNN we obtained were 300 hidden layer nodes, softplus for the input and output layers and relu for the hidden layer as activation function, adam optimizer, and batch size 10. The color differences of the trained BPNN model varied from 0.08 to 2.19, with an average ΔE of 0.63± 0.5, which was less than the clinically acceptable threshold of 2.65(p < 0.001).
Conclusions: The in-house developed BPNN model demonstrated the prediction of high-accuracy and can further help dental clinicians to determine appropriate colors for patients.
Methods: In this study, CIE L*a*b* values were retrieved from the shade guide (Vita 3D). To train the BPNN, each color shade tab was measured six times to form a dataset of 90 values. The training process randomly selected 63 values from the dataset as the training data and the remaining values as the testing data. We constructed the BPNN with Python and used the Keras framework. The BPNN architecture consisted of three input variables, two hidden layers, and four output variables. The parameters of the BPNN, including the number of hidden layer nodes, activation function, optimizer, and batch size, were determined by a systematic test. A tooth color was arbitrarily picked and passed through the trained BPNN to obtain the predicted output. Finally, the color differences(ΔE) between the real tooth color and the prediction were calculated.
Results: The optimal parameters of the BPNN we obtained were 300 hidden layer nodes, softplus for the input and output layers and relu for the hidden layer as activation function, adam optimizer, and batch size 10. The color differences of the trained BPNN model varied from 0.08 to 2.19, with an average ΔE of 0.63± 0.5, which was less than the clinically acceptable threshold of 2.65(p < 0.001).
Conclusions: The in-house developed BPNN model demonstrated the prediction of high-accuracy and can further help dental clinicians to determine appropriate colors for patients.