Early Recognition Model Of Patients With Bipolar Disorder Via Diffusion Tensor Imaging

Patients with bipolar disorder are at serious risk of being misdiagnosed as unipolar disorder in the early stage of the disease.Misdiagnosis can cause deleterious consequences for treatment.Currently,clinical diagnosis depends heavily on the subjective judgment of physicians.Advances in neuroimaging have made it possible to find objective biomarkers.It also provides adequate data for the construction of the classification model.Combining DTI with graph theory,previous studies have found great differences between patients with bipolar disorder and healthy people.However,more research is needed to explore the differences between the network of unipolar and bipolar disorder.Meanwhile,reliable classification models are still lacking for the identification of the patients with unipolar and bipolar disorder.A total of 62 patients with unipolar disorder and 112 patients with bipolar disorder are included in the study.Focusing on the different mechanism of unipolar and bipolar patients,this paper uses the brain structure network to study the topological attribute differences between the two diseases in the network.In addition,it constructs a classification model based on graph convolution network to distinguish the two diseases.The main contents are as follows:1.Investigate the differences of the rich club and topological properties in the brain structural network between patients with unipolar and bipolar disorder.Firstly,the brain structure networks of patients are constructed based on DTI data.After that,the rich club and topology properties of the brain network are calculated and compared statistically.At the same time,the rich club nodes and non-rich club nodes in this study are defined based on priori knowledge,in which the edges connecting rich club and non-rich club nodes are defined as feeder connections and the edges connecting between non-rich club nodes are defined as local connections.The findings suggest that patients with bipolar disorder have more white matter fiber bundle injuries than patients with unipolar disorder.The injuries are concentrated in feeder and local connections,which is consistent with the significant decrease in the global efficiency of the network in bipolar disorder.A decrease in the strength between feeder and local connections is found in the left hemisphere of the patients with bipolar disorder,indicating a decrease in the connection of non-hub nodes in the left hemisphere.A significant decrease in the node cluster coefficient and eigenvector centrality is found in the olfactory cortex,posterior cingulate gyrus and temporal lobe of bipolar patients,depicting that patients with bipolar disorder suffer more damage in the emotional,cognitive and reward-related connections than patients with unipolar disorder.2.Construct a graph convolution pooling model to classify brain networks of patients with unipolar and bipolar disorder.On one hand,the model avoids the loss of information caused by feature extraction.On the other hand,it preserves the transfering information among nodes.The results show that the graph convolution pooling model has excellent classification performance for the brain structure network,achieving an accuracy of 85.54% in the classification of patients with unipolar and bipolar disorder.The model further finds the difference between the left amygdala and insula in patients with unipolar and bipolar disorder,indicating impairment of emotion management for patients with bipolar disorder.Finally,the manifold dimension reduction method verifies the classification performance of the model.3.In order to make full use of the functional integration of human brain,a graph convolution ensemble model is constructed by combining the integration strategy of input attribute and training sample disturbance in ensemble learning.The brain network is divided into five hierarchical sub-networks based on the rich club property.In order to improve the divergence of the sub-classifier and balance the number of samples in the training set,the classification samples of each sub-classifier are constructed by a combined extraction method.The results show that the model achieves a higher accuracy of 87.74% and is more stable.The study of the sub-classifier confirms the difference of feeder and local connections between the two types of diseases.