Research On Computer-aided Diagnosis Of Alzheimer’s Disease Based On PET Images

The incidence and base of AD(Alzheimer’s disease)has increased as the global population has intensified and aging continues to increase at home and abroad,and AD is a far-reaching neurological disorder that poses a serious harm to patients and their families,as well as a great stress on society as a whole.Normal people go through MCI(Mild cognitive impairment)stage in the process of transitioning from healthy to AD.Therefore,it is important to accurately distinguish between AD,MCI and normal people.The diagnosis is not always accurate and the process is time consuming,as the doctor’s visual assessment can be limited by experience.This is where the involvement of a computer is needed.With the rapid development of science and technology,cutting-edge imaging technologies such as sMRI(Structural Magnetic Resonance Imaging),DTI(Diffusion Tensor Image),MRI(Magnetic Resonance Imaging)and PET(Positron Emission Tomography)have become important tools for identifying AD.PET is a molecular imaging technique with high sensitivity,high resolution and low loss,and is the highest level of radiation detection in radiological technology.With the rapid development of imaging technology,PET is gaining more and more attention as an important imaging-based technology with a very wide range of diagnostic applications,such as visualizing the morphological characteristics of Alzheimer’s lesions and their distribution status.Therefore,in order to contribute to a better application of computeraided diagnosis to improve diagnosis,the current state of research on computer-aided diagnosis based on PET images is studied in depth and imported into a deep learning network for training after pre-processing the data.The main work and results accomplished in this thesis are as follows:For past computerized deep learning-aided diagnosis based on PET images,optimization of the data itself is seldom found.While putting a lot of effort into the improvement of deep learning networks,some of the effort can be considered to be diverted to the optimization of PET data.The processing of PET images removes the distracting term like skull and segments the image into whole brain,white matter and gray matter.The best of the best is selected.For the EfficientNet-b0 network,this paper has made some improvements to the model.In order to save time and further improve the accuracy,several unimportant convolutional layers are removed and CBAM module and RBF module are inserted in this paper,which gives the network a considerable performance improvement.In cross-validation,the migration learning idea is introduced to load the parameters of the best previously saved model for iterative fine-tuning,which can shorten the training time as well as derive the optimal model.Translated with www.Deep L.com/Translator(free version)In the multi-model integration network,the fusion of Resnet50 network,ResNest18 network and DenseNet121 network is used to vote on the final results of the three unimodal results.NNI(Neural Network Intelligence,Microsoft)is inserted in it as a non-intrusive tool,in short,it is possible to implement automatic parametrization,automatic architecture search,etc.without changing almost anything in the original code.The score improvement strategy uses cosine annealing and integrated learning.