Research On Method Of Eye Movement Information Features Extraction And Vertigo Diagnosis Based On Neural Network

Vertigo is a common symptom with various causes,and it can be difficult to distinguish its causes.The symptoms of vertigo caused by different etiologies are often accompanied by different eye movement abnormalities.Therefore,this article starts research by mining the association between eye movement information and vertigo.Eye movement information includes conscious and unconscious changes such as pupil dilation,sight movement,saccade,and nystagmus.Many studies have proved that eye movement information can reflect the physical and psychological conditions of the human body to a certain extent,and some have even been applied in medical diagnosis,such as the Roll Test and Dix-Hallpike test,which can induce different types of nystagmus through changes in body position,and diagnose the semicircular canals of patients with benign paroxysmal positional vertigo(BPPV).This paper uses machine learning methods to distinguish between vertigo,brain damage and healthy control using eye videos in the visual tracking test.It also uses the eye videos in the rolling test to intelligently distinguish involved horizontal semicircular canals in BPPV.The main research contents of this paper are:(1)Vertigo,brain damage and healthy controls are classified according to the eye video in the target tracking experiments using neural networks.An eye tracker video acquisition device is made,and the supporting software is developed so that target tracking experiments could be presented.Three types of subjects,namely vertigo,brain injury,and healthy control are invited to take the target tracking experiments,and the eye tracker takes videos of the subjects’ eyes during the experiments.Applying the image processing method,the eye movement information is extracted to form feature matrixes.Then,a weak classifier is constructed for each feature based on a long short-term memory(LSTM)neural network combined with a fully connected network.Indicators are designed to quantify the classification ability of these weak classifiers.Finally,a strong classifier is constructed by the weighted average of the classification ability and the output result of the classifiers,so as to obtain the final classification result.(2)The neural network in(1)is optimized.For the weak classifiers in(1),a strong classifier is constructed in another way,using the random forest method.The output of the weak classifier is used as the input feature of the random forest,and the original three-classification problem is transformed into three two-classification problems.The construction of a single decision tree in the random forest follows the C4.5 construction method.(3)Different BPPV caused by the involvement of different semicircular canals was classified according to the eye video in the posture induction test using neural networks.The videos produced by Xinqiao Hospital during the diagnosis and treatment of BPPV patients are collected.The goal is to distinguish patients with canalolithiasis of the horizontal canal and patients with cupulolithiasis of the horizontal canal according to the nystagmus in the lying position,the left position,and the right position in the Roll Test.For the videos of these three positions,video segments of fixed length are cropped respectively.For each video segment,a three-dimensional matrix can be formed.The three dimensions are: the width of each frame,and the width of each frame,and the frames.Therefore,three three-dimensional matrices of the same size can be obtained from each person.For these three-dimensional matrices,three methods are applied directly for classification.The first is feed a three-dimensional convolutional neural network separately,then stack them up for another three-dimensional convolutional neural network.The second is to feed a LSTM network separately,then stack them up for another LSTM network.The third is to feed a two-dimensional convolutional neural network followed by a LSTM network separately,then stack them up for another LSTM network.Thus,three classifiers with certain classification capabilities are obtained,but don’t work well enough.So the second neural network model is optimized by combining optical flow and gradient to form multiple channels,which comprehensively utilizes the original video grayscale,the video grayscale change rate,and the motion trend in the video.The multi-channel neural network effectively enhances the ability to extract motion information and achieves better results.