Feature Detection And Classification Of Eye Movement Position Curve Based On Machine Learning

The incidence of vertigo is extremely high among middle-aged and elderly people,and there is a trend of increasing and younger.Eye movement is an important window to reflect the condition of vertigo.Therefore,eye movement test is the main diagnosis method of vertigo,and the result is often recorded as the eye movement position curve.The field of vertigo is relatively small,and the development of it’s medical facilities started late in China.At present,the clinic mainly relies on the doctor’s experience to analyze the eye movement position curve to make a diagnosis,which is relatively timeconsuming and laborious.Exploring the intelligent analysis method of eye movement position curve is of great significance for improving the efficiency and accuracy of vertigo diagnosis.The core components of eye movement in eye movement test are saccade and smooth pursuit,and the corresponding basic tests are saccade test and smooth pursuit test.The two tests measure the ability of patient’s eyes to pursuit the target quickly and slowly,which can provide an important reference for the diagnosis of vertigo to a certain extent.With the support of Science and Technology Commission of Shanghai Municipality and a vertigo equipment company in Shanghai,aiming at the intelligent analysis of eye movement position curve of saccade test and smooth pursuit test,this thesis mainly completes the following work:1.At present,there are no publicly available data sets of eye movement position curve.In this thesis,the most advanced clinical equipment is used to collect the eye movement position curve of saccade test and smooth pursuit test of more than 700 patients.According to the characteristics of the curves,four kinds of data marking software are designed and implemented,then the data marking is completed in cooperation with professional doctors.Finally,several small-scale data sets of saccade test and smooth pursuit test are constructed.2.In clinical practice,the eye movement position curve of the saccade test is mainly evaluated by identifying the undershot and overshoot of the curve,which is called the component analysis of the saccade test curve.In this thesis,the task is divided into two parts: saccade test curve segmentation and saccade curve segment classification.The former is achieved by building Uneye model based on U-Net,and the Cohen Kappa and F1 values reach 79.4% and 97.2%,respectively.The latter is done by building a MCNN model,and the F1 value reaches 95.8%.Finally,the Uneye+MCNN hybrid model is proposed to realize the component analysis of the saccade test curve.3.The eye movement position curve of the smooth pursuit test has a clear evaluation standard of four categories in clinical practice.Aiming at the problem of classification,the effects of traditional machine learning classifier and improved 1DCNN models are compared and analyzed.Finally,a hybrid model based on Stacking algorithm is proposed,which combines the advantages of the two types of models and achieves relatively good results in the test set of this thesis,with an F1 value of 87.3%.At present,there is a lack of related research on the intelligent analysis of clinical basic eye movement position curve.From the perspective of exploration,this thesis attempts and compares the effects of various machine learning methods including neural networks and traditional classifiers,and then obtains some relatively good models on the data sets of this thesis.The experimental results show that the methods proposed in this thesis can effectively analyze the eye movement position curve of saccade test and smooth pursuit test,and provide a certain reference value for clinical diagnosis.The work of this thesis also makes a certain contribution to gradually realizing the intelligent analysis of eye movement position test and the intelligent diagnosis of vertigo.