On construction of transfer learning for facial symmetry assessment before and after orthognathic surgery. (March 2021)

• Record Type: Journal Article
• Authors: Lin, Hsiu-Hsia; Chiang, Wen-Chung; Yang, Chao-Tung; Cheng, Chun-Tse; Zhang, Tianyi; Lo, Lun-Jou
• Is Part Of: Computer methods and programs in biomedicine. Volume 200(2021)
• Publication Date: 2021-03
• Subjects: Transfer learning -- CNN -- Facial symmetry -- Deep learning -- Data preprocessing; Medicine -- Computer programs -- Periodicals; Biology -- Computer programs -- Periodicals; Computers -- Periodicals; Medicine -- Periodicals; Médecine -- Logiciels -- Périodiques; Biologie -- Logiciels -- Périodiques; Biology -- Computer programs; Medicine -- Computer programs; Periodicals; Electronic journals; 610.28
• DOI: 10.1016/j.cmpb.2021.105928
• Languages: English
• ISSNs: 0169-2607

Abstract

Highlights: Convert 3D CBCT scan images of 71 subjects to 2D contour maps with contour map of 20 lines. Image Pre-processing based on contour map, including rotation, scaling, and clipping for data amplification. Implement Transfer Learning to train a new facial symmetry feature model. Predict facial symmetry scores for new data using the new trained model. Abstract: Orthognathic surgery (OGS) is frequently used to correct facial deformities associated with skeletal malocclusion and facial asymmetry. An accurate evaluation of facial symmetry is a critical for precise surgical planning and the execution of OGS. However, no facial symmetry scoring standard is available. Typically, orthodontists or physicians simply judge facial symmetry. Therefore, maintaining accuracy is difficult. We propose a convolutional neural network with a transfer learning approach for facial symmetry assessment based on 3-dimensional (3D) features to assist physicians in enhancing medical treatments. We trained a new model to score facial symmetry using transfer learning. Cone-beam computed tomography scans in 3D were transformed into contour maps that preserved 3D characteristics. We used various data preprocessing and amplification methods to determine the optimal results. The original data were enlarged by 100 times. We compared the quality of the four models in our experiment, and the neural network architecture was used in the analysis to import the pretraining model. We also increased the number of layers, and the classification layer was fully connected. We input random deformation data during training and dropout to prevent the model from overfitting. In our experimental results, the Xception model and the constant data amplification approach achieved an accuracy rate of 90%.