Research On Torsional Nystagmus Detection For Vertigo Diagnosis

Dizziness and vertigo are the common clinical symptoms and typical symptoms of many diseases.In China’s large population and aging society,the number of patients with dizziness and vertigo is increasing.Nystagmus is the most sensitive and specific sign of vestibular lesions in the bedside examination of dizziness and vertigo.The measurement of nystagmus pattern by eye movement video collected in clinic can provide valuable diagnostic basis for dizziness and vertigo.However,the above process still depends on experts and specialist examination,which cannot realize intelligent nystagmus pattern recognition and bedside diagnosis.Benign paroxysmal positional vertigo(BPPV)is a very common vertigo disease,and torsional nystagmus is its main sign.Therefore,for the torsional nystagmus with a high incidence rate,excavate its nystagmus characteristics,establish a torsional nystagmus detection model based on deep learning,realize automatic nystagmus detection and interval positioning,which can effectively relieve the doctor’s work pressure and improve its diagnosis efficiency and provide support for the final auxiliary diagnostic system.The main work and contributions of this thesis include:(1)To solve the problem of the positioning of the pupil center under the interference of eyelashes,eyelids,etc.,a method for nystagmus video condensation and calibration is proposed.Using convolutional neural network to eliminate invalid frames in nystagmus video,which avoids the interference of occlusion on the positioning of the pupil center,also reduces the workload of doctors to browse nystagmus videos.On this basis,to solve the problem of the horizontal and vertical displacement of the eyeball disturbing the analysis of torsional motion,a pupil registration method(CHt-TM)based on template matching based Hough transform circle detection and trajectory tracking is further proposed.The problems of eyelash interference and pupil deformation are combined with template matching to complete pupil registration based on time and space consistency.Experiments verify the performance of nystagmus video condensation and calibration methods.(2)In order to solve the problem of automatic detection of weak torsional nystagmus,a detection method of torsional nystagmus based on deep learning is proposed.Using optical flow to estimate the trajectory of nystagmus,and establishing a torsion aware motion confidence map for visualizing the movement trend and intensity;on this basis,training a torsion aware two path way convolutional neural network to complete the identification of nystagmus images;finally introducing error correction rules based on time and space consistency to achieve the merging of nystagmus images to complete the location of the final torsional nystagmus segment.Experiments are carried out on clinically collected eye movement video dataset,and the experimental results verify the effectiveness of the scheme.(3)In order to improve the efficiency of the labeling of torsional nystagmus in nystagmus video,an intelligent assisted eye tracking video annotation system is constructed.On the Node.js platform,using the Vue.js front-end development framework and Webpack build tools,building a annotation system and deploying it to the cloud server.The system can realize video playback,control,mode selection,clip annotation and other functions.The experimental data annotation files in this thesis are collected by this system,which can verifies the stability and availability of the system.