Early Diagnosis Of Alzheimer's Disease Based On Deep Learning

Alzheimer's disease is a neurodegenerative disease that usually occurs in the elderly,also known as Alzheimer's disease.According to the difference of degrees of this disease,it can be divided into three phases: Alzheimer's disease(AD),mild cognitive impairment(MCI)and normal control(NC).AD is an incurable disease.If we can diagnosis the disease early and take timely solution to treat and maintain,we can delay the deterioration of this disease and reduce the impact on patients and their families.Therefore,the early diagnosis of Alzheimer's disease is of great significance.In this paper,the magnetic resonance imaging(MRI)from ADNI is used on the early diagnosis of AD based on deep learning.The specific research contents are as follow:(1)The method that trains the convolutional neural network(CNN)from end to end based on fine tuning is proposed on the early diagnosis of AD.There are two stages on the early diagnosis of AD based on traditional machine learning methods including features extraction based on the prior knowledge and classifier constructed.Features extraction and classifier training are separated during diagnosis.In this paper,the method that trains the convolutional neural network from end to end by fine tuning is used on the early diagnosis of AD.Convolution neural network is composed of convolutional basis and classifier.Firstly,the weight of convolutional base is not changed,we only change the weight of classifier during the training.When the model performance tends to be stable,we unfrozen the top convolutional layers of convolutional base.Then,we train the trained classifier combined with the convolutional base from end to end to learn the high-level features on the diagnosis of AD.In order to improve the performance of the results,ensemble learning is introduced into this paper.Res Net50,NASNet and Mobile Net are selected as the base classifiers.In the last,the results from the base classifiers are ensembled as the last results for the diagnosis of AD.The accuracy from AD and MCI is up to 97.65%.(2)Selective kernel network with attention mechanism(SKANet)is proposed on the early diagnosis of AD.Generally,CNN can increase the depth of the network by stacking the convolutional layers.Different from ordinary CNN,the depth of SKANet is increased by stacking the residual blocks with the same topology for saving the complexity of model design.Selective kernel(SK)convolution is introduced into the residual blocks.Generally,ordinary convolutional layer is designed by single size of the kernel in the same layer to extract features.SK convolutional layer is designed by different kernel sizes to adaptively adjust the receptive field by gate mechanism in the same layer.Then,attention mechanism is added to the bottom of the block to emphasize on important features and suppress unnecessary ones for more accurate representation of the network.The accuracy on 3-way classification from AD,MCI and NC is up to 94.21% of SKANet.The accuracy from AD and MCI is up to 98.82%.The results confirmed the validity of SKANet on the early diagnosis of AD.