Research On Classification Diagnosis Of Depression Based On Graph Convolutional Neural Network

Depression is a common mental illness and frequently observed in the fast-paced and intensely competitive social environment.It is essential to diagnose and subdivide serious clinical depression,and then adopt suitable method of treatment to the case.However,the current clinical diagnosis of depression is mainly by asking the patient’s condition and then judging according to the criteria.Such a diagnosis is relatively covert and affected by the doctor’s experience and level,so the diagnosis result is subjective and prone to misdiagnosis.In recent years,many studies have sought for biological diagnostic indicators of depression from the perspective of brain medical imaging.The commonly used methods directly adopt functional connection weights of different brain regions as the feature for learning and analysis,but this feature ignores the topological structure information of brain networks.Therefore,the diagnosis accuracy is not high.This paper proposes a classification method for depression with high accuracy,which is designed to identify whether a person suffered depression(Major Depressive Disorder,MDD)or not(Healthy Control,HC).If so,it then designed whether the depressed patients can be effectively treated by antidepressant therapy(Responsive Depression,RD)or not(NonResponsive Depression,NRD).This paper implements two innovations in model method and disease diagnosis.In terms of model method,the brain graph extracted from functional Magnetic Resonance Imaging(fMRI)is innovatively used as the diagnostic basis,which describes the topology of functional connectivity between brain regions.And then this data will be sent into the Graph Convolutional Neural Network(Graph-CNN)for feature extraction and depression diagnosis classification.In terms of disease diagnosis,this thesis innovatively studies the distinction between RD and NRD to provide appropriate therapy for each patient with MDD,i.e.,predicting efficacy of antidepressant therapy.This thesis will expound from three aspects: data preparation,design of diagnostic system and experimental verification.Firstly,data acquisition and preprocessing are required with the type of fMRI.And then construct the brain graph which contains four steps:(1)extraction of BOLD signal of brain regions,(2)generation of brain function network connection matrix,(3)binarization of the adjacency matrix and(4)construction of brain graph described topological structure.Secondly,the constructed brain graph and the actual diagnostic label were input into Graph-CNN for feature learning and model training,where the feature analysis and extraction of brain map data is realized through the combined operation of graph convolution and graph prior pooling.The prior pooling of the graph includes two methods: Similar Indegree Pooling(SIP)or Whole Brain Hierarchical Network Pooling(BHP).Next,the final disease classification and efficacy prediction are achieved through the fully connected layer.Thirdly,5-fold cross-validation is adopted to verify the accuracy of the result of diagnosis,finding that the accuracy of disease diagnosis(HC or MDD)can reach up to 90.11% through the GraphCNN-SIP under the map of Anatomical Automatic Labeling(AAL)atlas and threshold quantization(T = 0.8),and the accuracy of three classification(HC,RD or NRD)is up to 79.12%.The accuracy of the efficacy prediction(RD or NRD)is up to 80.44% through GraphCNN-SIP under the map of Brainnetome(BN)atlas and threshold quantization(T = 0.8).Further study on the group of depressed patients,a rule can also be found between RDs and NRDs.It shows that the connectivity distribution of RDs is more intensive,while NRDs is sparser with baseline reference of HCs.Thus,this study explains the pathological mechanism of depression from the topological features of brain neurological function connection and laid a foundation for the systematic diagnosis of depression.Besides,the filter coefficients of the first convolutional layer in our network were extracted.And the best distinguished brain area,Left Anterior Cingulated Cortex,is found out and relevant medical verification and interpretation are performed to verify the validity and feasibility of this system.