Research On The Detection And Classification Of Flash Visual Evoked Potential Based On Deep Learning

As the most important organ of human,the eye is responsible for collecting visual information and transmitting it to the central nervous system.If the optic nerve pathway is damaged,it will affect the transmission of visual information,and may even cause blindness.Flash visual evoked potential(FVEP)is one of the commonly used techniques for diagnosing the integrity of the optic nerve pathway in clinical practice,and it can provide an important reference for clinicians in diagnosis and treatment.Doctors diagnose diseases based on FVEP,and need the peak time of the three positive waves and the valley time of the three negative waves,especially the peak time of the second positive wave P2.These six time points are the six feature points of FVEP,which need to be marked by a professional doctor.Due to the complexity of FVEP,labeling is difficult.Therefore,an accurate automatic detection system of feature points of FVEP increases work efficiency and reduces the misdiagnosis rate for doctors.Retinitis pigmentosa(RP)is a serious genetic disease that damages the optic nerve pathway and causes high blindness.The FVEP signal contains a wealth of information and is one of the tools for diagnosing RP.Therefore,FVEP can be automatically classified to diagnose RP.Based on deep learning,machine learning and probability theory,this thesis proposes efficient algorithms around feature points detection of FVEP and RP disease diagnosis.The main research contents of this thesis are summarized as follows:(1)An automatic detection framework of FVEP feature points based neural network and Light GBM is proposed.First,in response to the complex changes in FVEP,based on the convolutional neural network and attention mechanism,a neural network CAA-Net is designed to select a set of possible feature points from the complex FVEP signal.Next,a set of candidate sequences of feature points is generated,and K candidate sequences of feature points are selected using a multivariate Gaussian model.Then,using the Light GBM model,combined with the features of FVEP,select the optimal sequence of feature points.Finally,the clinically collected FVEP data set is used to verify the validity of the model,and the sequence mean absolute error is 9.79.(2)A novel multi-input neural network RP recognition and out-of-distribution detection framework based on convolution and attention mechanism is proposed.Aiming at the features of FVEP signal that have different information locally and globally,CSANet with global and local feature extraction is designed.For the detection of out-ofdistribution samples,a confidence branch is added to CSA-Net.For the proposed manual features,a new input layer is added to CSA-Net,and MCSA-Net is constructed from this.Finally,the clinically collected FVEP data set is to verify the effectiveness of the model.In the classification task,the accuracy is 98.7%,and in the out of distribution detection task,the accuracy is 95.5%.