Application Of Long-term And Short-term Memory Networks In The Diagnosis Of Brain Diseases

In recent years,brain disease is one of the most important causes of global population death.Neuroimaging plays an increasingly important role in clinical diagnosis and brain disease analysis.However,clinicians often need to spend a lot of time to do a diagnosis by it,because of the high dimension of neuroimaging data.Therefore,it is necessary to assist doctors in disease analysis and diagnosis by computer aided diagnosis.At present,some automatic diagnosis methods have been used in the analysis and automatic diagnosis of brain diseases.The results show that they can effectively analyze n neuroimaging and make diagnosis of diseases,but most of methods base on traditional machine learning or convolution neural network methods.However,neural images often have temporal information and feature maps have unique spatial information.According to the above analysis,we use this paper designed the diagnosis models based on long short-term memory(LSTM)in Alzheimer’s disease(AD)and obsessive-compulsive disorder(OCD).First,this paper replace the traditional LSTM by stacking two layers of LSTM.In addition,each output of the LSTM is related to the historical input,but has a greater impact on the current input.Therefore,the full connection layer can improve the accuracy by enhancing the connection between different output nodes from the stack bidirectional LSTM.Based on this observation,we designed a new deep learning network,which uses multimodal data to diagnosis AD through 3D neural roll and network and full stack two layers of LSTM.Finally,we have achieved 94.82% of the diagnostic accuracy.Secondly,we designed a new deep learning framework for OCD diagnosis based on functional magnetic resonance imaging(fMRI)data.Firstly,two independent time series are extracted from the original resting fMRI sequence by frame separation and sampling,which reduces the length of the resting fMRI sequence and reduces the training difficulty.Secondly,two kinds of independent bidirectional attention stacking long short-term memory(BAS-LSTM)are used to learn the hidden spatial information,and the preliminary classification results are obtained.Finally,two diagnostic results were voted to get the final diagnosis results.Finally,71.66% of the diagnostic accuracy was achieved.Thirdly,based on the work of the second part,we design a new framework of OCD diagnosis based on ordered neurons LSTM(ON-LSTM)and multi head attention.ON-LSTM can enhance the performance of LSTM by using the formation order of neurons instead of using superposition and bidirectional operation.We use the multi head attention method to extract various attention features from the ON-LSTM feature map,and extract information from different representation subspaces.Finally,we have achieved 68.89% classification accuracy.