Prediction And Analysis Of Multisource Heterogeneous Fundus Data Of Alzheimer

In recent years,the research on eye images has been a research hotspot in the medical field.For example,the segmentation of retinal blood vessels,the numerical extraction of blood vessel distribution angles in fundus images,and the cropping of vein vortices in fundus ultra-wide-angle images are all hot topics at this stage.In the past,the diagnosis of fundus images required a large amount of clinically collected patient data,and then an experienced ophthalmologist completed the diagnosis manually.There has been some progress in medical data prediction.However,fundus data has the problems of few datasets,different sizes of blood vessels,and discrete data,which also increases the difficulty of deep learning algorithms in predicting missing medical data,and also leads to accurate prediction of missing parts of inpatient medical data.Padding and accurate classification predictions are fraught with challenges.In summary,there is still a lot of room for improvement in the prediction of medical data.This paper selects a dataset（Alzheimer’s Disease Dataset,ADD）collected from the Institute of Biomedical Engineering,Peking Union Medical College,which contains Alzheimer’s Disease（AD）data and healthy person data.Read a lot of literature,choose Simple Shallow-Recurrent Neural Network（S-RNN）,Robust Auxiliary Classifier Generative Adversarial Network（r ACGAN）,Multi-layer Output U-Network（MO-Net）,and GG-Very-Deep-16 Convolutional Neural Network（VGG-16）and other networks are used as basic models,and improvements are made based on mastering these models.Compared with other existing methods,this method has a good effect on AD fundus data prediction.The main contributions of this paper mainly include:（1）Create an ADD dataset,compare the data used for AD patient diagnosis through meta-analysis,and determine the characteristics that need to be collected.The ADD dataset contains a total of 4195 examiners’ fundus data,including 2135 healthy people’s fundus image data and 2060 AD patients’ fundus image data.（2）The MO-Net network is used to crop the optic disk part of the conventional corner fundus image data in the ADD dataset to remove redundant information.In the experiment,the method of the U-Net network for semantic segmentation is improved,the original U-Net network is improved,and a side output layer is added to it,the position of the video network disk is located,and the regular angle image is cropped.The final output of the experiment is an image of the optic disc and its surrounding parts with size（800,800,3）.（3）Use the S-RNN network to fill in the important value of the Mini-Mental State Examination（MMSE）in the fundus data.The experiment improved the S-RNN network,fully understood the relationship between the input layer,the hidden layer,and the output layer in its network structure,and then collected the characteristics of the text data in the ADD data set,and used the existing complete text data to analyze the S-RNN.The network is trained to complete the MMSE numerical regression prediction filling for patients with unknown conditions.The final fit of the whole experiment is 76.07%.（4）Classification of multi-source heterogeneous AD fundus data using multivariate construction network.In the experiment,the r AC-GAN network and the VGG-16 network are used to classify the AD text data and image data respectively,output the classification prediction distribution of the two,and then use the Random Forest network to reclassify the classification prediction distribution of the two.The final results are that the VGG-16 image classification accuracy is 95.47%,the r AC-GAN image classification accuracy is 96.88% and the final classification accuracy is 99.39%.