3D Transfer Learning Network For Classification Of Alzheimer’s Disease With MRI

Alzheimer’s disease(AD)is a neurodegenerative senile dementia.Once diagnosed,it cannot be cured.As the country with the largest number of AD patients in the world,people are suffering from the pain and economic pressure caused by this disease,and the symptoms of the previous disease will be mistaken for natural aging,resulting in missing the best treatment time,so early diagnosis and medicine to delay the deterioration of the disease is very important.In recent years,with the rapid development of medical imaging technology and computer technology,the methods of machine learning and deep learning have received extensive attention in the field of neuroimaging.At present,most of the neuroimaging data that can be obtained by open source are complex high-dimensional data,so it is worth studying to effectively preprocess the data,retain enough valuable features,and select the network with low complexity to quickly obtain high classification accuracy.Existing classification methods have the advantages of high accuracy in AD classification,but there are still some problems such as limited available labeled medical data,high computational complexity of 3D convolutional network,cumbersome feature extraction process,difficulty in obtaining pre-training weights of 3D migration network and Alex Net network and the classification does not conform to medical practice.In view of this,this paper uses machine learning as an auxiliary diagnosis method for AD,and proposes a 3D transfer network based on 2D transfer network to classify AD and normal groups in Magnetic Resonance Imaging(MRI).Firstly,the 3D MRI data are sliced in different directions using the SPM12 toolkit,and appropriate slices are selected in different ways to form training and validation sets.Secondly,the 2D transfer network is used to extract features from the 2D slices of MRI,and the extracted features are further reduced in dimension.Finally,all 2D slice features of a subject are merged for classification.The experiment in this paper uses an open access Alzheimer’s disease database to evaluate the method.The experiment result show that the classification accuracy of the proposed 3D network is better than that of the existing 2D transfer network,increased by about 10 percentage points and the classification time is only about 1/4 of the existing one.And four pre-training networks,VGG16,Dens Net121,Inception V3 and Mobile Net were transferred in the proposed classification network bottleneck layer.The results show that the migration of Mobile Net network is more suitable for our method.In addition,in order to verify the feasibility and effectiveness of the algorithm in this paper,we use the ADNI dataset to test our method again,and the results are consistent.The method proposed in this paper realizes the classification of 3D MRI data through the existing 2D transfer network,which not only reduces the complexity of the traditional 3D network,but also improves the classification accuracy.The classification time is saved due to the weight sharing of the transfer network.