| Alzheimer’s Disease(AD),commonly known as senile dementia,is an irreversible and fatal chronic neurological Disease.Alzheimer’s disease currently affects 50 million people worldwide,and the number is increasing year by year.The course of this disease is slow,and there is no clinical plan that can completely cure this disease.What we can do is to reduce the symptoms and delay the disease through some drugs or psychological intervention in the precursor of this disease.Therefore,it is very important to diagnose the disease and its prodromal stage through medical imaging technology,which has great clinical significance for the follow-up treatment of patients.Functional Magnetic Resonance Imaging(f MRI),as a non-invasive,high-resolution neuroimaging technology,has been introduced into the diagnosis of Alzheimer’s disease with excellent performance.With the continuous development of artificial intelligence,machine learning has become a research hotspot.As a major branch of machine learning,deep learning has been widely used in the field of medical images.Therefore,combined with f MRI images and based on deep learning method,this paper carried out related classification prediction for the three stages of Alzheimer’s disease(AD,MCI and Normal Cognize(NC)).The main contents of this paper are as follows:(1)The current f MRI-based classification methods mostly use Function Connection(FC)as input,ignoring the dynamic time process.In contrast,deep learning models have excellent time representation ability,in which Recurrent Neural Network(RNN)can capture features from dynamic time.In this paper,a MCGRU model is built,which combines a Convolutional Neural Network(CNN)with a gated Recurrent Unit(GRU)in RNN.The accuracy was 92.3%,86.7% and 84.3% in the classification tasks of ADvs.NC,ADvs.MCI and MCIvs.NC.Finally,we set up several comparative experiments,the results show that the classification performance of this model is better than other comparative experiments.(2)Due to the high dimension of f MRI scan,the manual extraction of features will inevitably lead to the loss of valuable information.Secondly,f MRI data is a four-dimensional image,which may cause partial information loss when the four-dimensional image is reduced to a two-dimensional or three-dimensional image.Therefore,in order to fully learn the spatial and temporal information in f MRI images,a 3DCNN-LSTM model was built in this paper,which mainly combines the 3D convolutional neural network with the Long Short Term Memory(LSTM)network in the cyclic neural network.The model can directly process 4D f MRI data,that is,extract spatial features from 3D static images in f MRI data,and then map the obtained feature images into LSTM network to capture the time-varying information in the data.The accuracy of this model in three groups of classification tasks: ADvs.NC,ADvs.MCI and MCIvs.NC reached 93.8%,85.2% and 82.6%.In addition,compared with 3DCNN,2DCNN and traditional methods,the energy separation performance of 3DCNN-LSTM model is better than those of them.Finally,the influence of the structure and parameters of LSTM on the classification accuracy of the model is analyzed. |
