[Application of sample-efficient neural network in predicting postoperative facial profiles in skeletal class Ⅱ orthognathic patients].

OBJECTIVE: Soft tissue prediction is of great importance for the surgical design and patients ' satisfaction in orthognathic surgery. Previous studies either rely on elastic mechanical models, such as finite element which involves time-consuming computations, or using artificial intelligence method which require insatiable large patient data. Therefore, in this study, we proposed a more efficient and lightweight method for facial morphology prediction and validated its accuracy. METHODS: A total of 85 skeletal class Ⅱ patients who underwent double jaw surgery performed by the same senior surgeon were included in the study. Three-dimensional (3D) facial images were acquired before and 6 months after surgery. Among them, 72 data sets were utilized to train the neural network. The Basel face model (BFM), a 3D morphable face model, was employed to parameterically reduce the dimensionality of the three-dimensional facial data. An artificial neural network (ANN) was trained to learn the mapping relationship between the preoperative and postoperative latent codes. The predicted postoperative latent code was subsequently decoded through BFM to reconstruct the three-dimensional facial morphology and gene-rate the predicted postoperative facial appearance. Thirteen patients excluded from the training set were used as the test set. The postoperative facial soft tissue morphology obtained from each patient was considered as the ground truth. The postoperative facial predictions generated by proposed neural network model were defined as the experimental group, while predictions generated by the orthognathic surgical planning software ProPlan CMF were used as the control group. The prediction accuracy was quantitatively evaluated by calculating the root mean square error (RMSE), mean absolute error (MAE), and the percentage of points with prediction errors within ±1 mm and ±2 mm, namely the inlier overlap percen-tages [IO(±1 mm)% and IO(±2 mm)%]. RESULTS: The experimental group exhibited a mean RMSE of (0.96±0.15) mm and MAE of (0.84±0.12) mm for the entire face. The IO(±1 mm)% and IO(±2 mm)% were (72.17±9.87)% and (96.85±2.60)%, respectively. The errors in the chin and lower lip regions were relatively larger than in other facial regions, with RMSE values of (1.35±0.42) mm and (1.47±0.46) mm, and all results were under 2 mm (P < 0.05). For the control group, the RMSE was (1.69±0.47) mm, MAE was (1.37±0.39) mm, IO(±1 mm)% and IO(±2 mm)% were (51.31±14.39)% and (79.69±13.71)%, respectively. The experimental group demonstrated smaller RMSE and MAE for the entire face compared with the control group (P < 0.001), and higher IO(±1 mm)% and IO(±2 mm)% (P < 0.001). In each region, there was no statistically significant difference in the RMSE between the experimental and control groups in the upper lip and lower lip regions. However, in the other regions, the RMSE in the experimental group was significantly smaller than in the control group (P < 0.05). CONCLUSION: The accuracy of a soft-tissue prediction system using neural network system based on BFM is clinically feasible.