Research On FMRI Classification Based On LSTM And GCN

Attention Deficit Hyperactivity Disorder(Attention Deficit Hyperactivity Disorder,ADHD)is also known as ADHD,as the name implies,patients with the disorder will appear the symptom of "Hyperactivity".The symptoms include too exuberant energy,difficulty concentrating,and impulse the phenomenon such as personality.The disease is more common in boys,and recent experiments have proved that it could continue into adulthood.Doctors usually diagnose this disorder by comparing the clinical symptoms with the international diagnostic criteria,which is relatively subjective.So it has aroused people’s attention to find a quantitative technology that is helpful for the diagnosis of ADHD.In recent years,with the emergence of functional Magnetic Resonance Image(fMRI),which is non-invasive,non-radiation,with high temporal and spatial resolution and does not require the subjects to complete complex tasks,it has been widely used in the study of mental diseases.Recently,deep learning technology has attracted increasing attention in the field of neuroimaging.Papers from the resting state fMRI data of ADHD current research progress and significance,aiming at the shortcomings in the course of the research process of ADHD,introducing deep learning related research,combined with both short-term and long-term Memory Network((Long-Short Term Memory,LSTM)),figure convolution Neural networks(Graph Convolutional Neural Network,GCN)and online dictionary learning,realize the classification of ADHD patients and normal control group.The final results can provide a method for neuropsychiatrists to accurately diagnose ADHD patients and improve the diagnostic efficiency.The main research contents of this paper are as follows:(1)Starting from the current clinical diagnosis of ADHD at home and abroad and the development of fMRI data,this paper introduces some methods to process fMRI data through machine learning and deep learning and apply it to the diagnosis classification of ADHD,and gives a theoretical introduction to each corresponding algorithm and network model.(2)For the processing Of resting state fMRI data,first Of all,the online dictionary learning method was used to extract the Region Of Interest(ROI)Of the brain;then,in order to achieve better learning effect Of the model,the data was expanded by means Of time domain transformation;finally,the classification Of patients and normal people was realized by LSTM.The accuracy,sensitivity and specificity of the classification results were 79.01%,62.70% and 88.90%.The effectiveness of the proposed method for the classification of ADHD was verified by experiments.(3)As fMRI is a four-dimensional image,in order to make full use of its spatial features,this paper introduces the graphic convolutional neural network.GCN can effectively extract the spatial image features of fMRI data,combine with LSTM,and finally classify them.In order to better extract timing series features with subsequent LSTM,we used Dynamic Functional Connectivity Matrix as the input of GCN model,and finally obtained classification results with an accuracy of 80.20%,sensitivity of 63.5% and specificity of 89.0%,which was 1.19% higher than the previous method that only used LSTM to process ROI time series.The results show that the combination of GCN and LSTM model is helpful to the classification of ADHD,and provides a possibility for the direction of subsequent clinical diagnosis.