Study On Facial Nerve Functional Assessment System Based On Artificial Intelligence

Objective:For developing a facial nerve functional assessment system based on artificial intelligence(AI),a database of facial images for patients with oral and maxillofacial diseases(OMD)was set up.This database was used to evaluate and improve the accuracy of an existing algorithm for facial-landmark detection.On the basis of facial-landmark detection,diagnostic indicators were measured and screened,then finally used to calculate modified regional facial nerve index(RFNI)and facial nerve index(FNI).The AI was trained for assessing patients with facial palsy according to House-Brackmann(HB)grading system and Facial Nerve Grading System 2.0(FNGS 2.0),to establish a method for facial nerve functional assessment.Ultimately,completing a mobile application(APP)for facial nerve functional assessment and individual guidance of functional training,and making it possible to clinical application.Methods:Images of 8 facial movements from OMD patients were collected.All images were labelled as normal or abnormal according to their facial features,and landmarks were detected by a convolutional neural network called HRNet.10 fold cross-validation was executed to evaluated the accuracy,which was measured by normalized mean error of testing sets before and after training.Images from healthy people and patients with facial palsy were collected.After detected by HRNet,objective indicators were measured.Delphi method was employed to get HB grade and FNGS 2.0 score as subjective diagnosis.Indicators were screened according to its correlation with subjective diagnosis,and then screened indicators were sum up as RFNI and FNI.Comparing correlation between RFNI and FNGS 2.0,FNI and HB grade.Using unscreened indicators and subjective diagnosis as input data to train Softmax.10 fold cross-validation was executed to train and get output diagnosis.Comparing output diagnosis and subjective diagnosis to calculate accuracy of the model.The facial nerve functional assessment system based on AI was programmed,in which diagnosis and guidance of functional training were involved.The guidance videos can be allocated individually according to the diagnosis of patientsResults:912 images from 300 OMD patients were collected.Accuracy for the abnormal group was lower than that for the normal group before training,but improvements in accuracy were identified in both groups post-training.In total 193 sets of images were collected from 151 subjects.Diagnostic indicators originated from different regions were screened and used for calculating RFNI and FNI.Statistically correlations were demonstrated between RFNI and FNGS 2.0 score,FNI and HB grade.The accuracy of Softmax was between 30.41%-41.37%.The APP for facial nerve functional assessment and individual guidance of functional training was completed preliminarily.It involved 4 modules:diagnosis,training,statistic and feedback.With expansion of the database,HRNet and Softmax could be retrained to reach a better accuracy.Conclusion:In this study,a preliminary OMD database was set up,and the accuracy of HRNet in facial-landmark detection was significantly improved after training with this database.Based on the detection of facial landmarks,diagnostic indicators were screened and a modifying method for calculating RFNI was established.A method for facial nerve functional assessment based on AI was also established,with accuracy yet to be further improved by expanding the training database.The APP for facial nerve functional assessment and individual guidance of functional training was completed preliminarily,with functions need to be further improved.