| Systemic lupus erythematosus(SLE)is a typical autoimmune disease,and neuropsychiatric systemic lupus erythematosus(NPSLE)is a common and serious complication in SLE patients,but its pathogenesis has not yet been elucidated.The clinical symptoms of non-neuropsychiatric SLE(non-NPSLE)are complex,insidious,and lack of objective indicators for diagnosis.As the disease progresses,if it develops into NPSLE,it may indicate poor prognosis or increase in treatment difficulty.There is an urgent need to develop clinical practical technical means and objective indicators to make early judgments or interventions for patients to improve prognosis.Magnetic resonance imaging(MRI)technology and brain network analysis methods are important means to study brain damage in complex diseases such as SLE.Machine learning is a bridge between basic research and clinical applications of neuroimaging,and has great potential in clinical applications.This study took SLE patients as the research object and used multimodal MRI as a tool to explore the characteristics and neuropathological mechanisms of brain damage in SLE patients from the levels of brain microstructure,brain structure network,and functional connectivity network.And use a variety of MRI imaging indicators to conduct machine learning research,explore the application value of MRI-based neuroimaging features in SLE classification,and find potential imaging biomarkers of SLE.The research content of this study is divided into the following four parts:Part Ⅰ:Study on Cerebral Cortex Volume,Surface Area and Their Structural Covariance Networks in Patients with Systemic Lupus ErythematosusObjectives:Using structural magnetic resonance imaging(sMRI)data to study the characteristics of cerebral cortical microstructure and structural covariance networks(SCNs)in SLE patients.Methods:The cortical surface area and cortical volume of 127 SLE patients and 102 healthy controls were calculated according to the surface-based morphology(SBM)method using FreeSurfer software,and the general linear model(GLM)was used to compare the cortical surface area and cortical volume between the two groups,respectively(results were corrected for multiple comparisons using Monte Carlo simulations with 10,000 permutations,cluster-forming threshold(CFT)P<0.001);Spearman correlation analysis was performed to compare the Systemic lupus erythematosus disease activity Index(SLEDAI)、Mini-mental State Examination(MMSE),Hamilton Anxiety Scale(HAMA),Hamilton Depression Scale(HAMD)of SLE patients with the cortical surface area and cortical volume of the brain regions with differences between the groups(P<0.05).According to Desikan Killiany atlas,the mean cortical surface area and mean cortical volume of 68 brain regions were extracted from all subjects.The differences in network global properties,local properties and network hubs of the two groups of cortical surface area and volume SCNs were constructed and compared using Graph Analysis Toolbox(GAT),respectively(FDR correction,q<0.05,two-tailed).Results:There was no significant difference in age,sex composition and cortical surface area between SLE group and HCs group.Compared with the HCs group,the cortical volumes of a total of 17 clusters in the SLE group,including the left superior frontal gyrus,left paracentral lobule,right precentral gyrus,and bilateral superior frontal gyrus,were significantly reduced.The SLEDAI was negatively correlated with the volume of the right parahippocampal gyrus,the MMSE total score was positively correlated with the volume of the right parahippocampal gyrus,and the HAMA total score was correlated with the volumes of 9 brain regions including the bilateral superior frontal gyrus and the right parahippocampal gyrus,respectively.The total score of HAMD was positively correlated with the volume of 8 brain regions including the bilateral superior frontal gyrus,and it was highly overlapped with the brain regions related to HAMA and HAMD.SCNs analysis showed that the normalized clustering coefficient(y),global efficiency(Eglob),modularity of SLE group was significantly lower than that of HCs group.In the four local attributes,including the local clustering coefficient,betweenness centrality,degree centrality and local efficiency(Eloc),the brain regions with increased and decreased indexes coexisted.In the SLE group,SCNs had four hubs in left middletemporal,left rostralmiddlefrontal,left superiortemporal and right parahippocampal,which were different from HCs in number and location.The analysis of cortical volume SCNs found that there was no significant difference in the global indicators between the two groups;4 local indicators also showed the coexistence of brain regions with increased and decreased indicators;The SCNs in the cortical volume of the SLE group had two hubs,the orbital part of the left parsorbitalis and the left rostralmiddlefrontal,which were also different from HCs in number and location.Conclusions:Cortical microstructure and its structural covariation network are damaged in SLE patients.The reduction of cortical volume in some brain regions may be related to disease activity,cognitive impairment,anxiety and depression in SLE patients.These changes may be the brain structural basis of neuropsychiatric symptoms in SLE patients,and anxiety and depression in SLE may share similar or even the same neuropathological pathways.Part Ⅱ:Study on the microstructure and structural network of cerebral white matter in patients with systemic lupus erythematosusObjectives:Using diffusion tensor imaging(DTI)data to study the characteristics of white matter microstructure and white matter structural network changes in SLE patients.Methods:The PANDA software was used to preprocess the DTI data and build the FA image and average FA skeleton.TBSS analysis was performed using FSL software to compare FA values between the two groups.The 116 brain regions divided according to AAL-116 atlas were used as nodes of the white matter structure network.Fiber Assignment by Continuous Tracking(FACT)method was used for deterministic tracking to reconstruct the whole white matter fiber bundles of each subject and calculate the FA value and fiber number of each node pair defined above.FN).Gretna software was used to analyze the white matter structure network based on graph theory,and the differences between the global and local attributes of the two groups of white matter structure network were compared(FDR correction,q<0.05).Spearman correlation analysis was performed on the clinical data and FA values,network indexes of the brain regions with differences between groups.Results:TBSS-based intergroup comparisons showed that FA values of five white matter clusters in SLE patients were significantly reduced compared with those in HCs,and no cluster with significantly higher FA values were found in SLE patients compared with HCs.SLEDAI was negatively correlated with FA values in the knee of the corpus callosum,and HAMD was negatively correlated with FA values in the splenium of corpus callosum and retrolenticular part of internal capsule.In the global network attributes,Cp,Eglob,Eloc was significantly decreased and Lp was significantly increased in SLE group compared with HCs group.In terms of local attributes,the node degree in SLE group was significantly decreased in 6 brain regions including the right superior temporal gyrus,and the node efficiency in SLE group was significantly decreased in 17 brain regions including the right thalamus.There was no correlation between SLEDAI,MMSE,HAMD,HAMA and global attributes;SLEDAI was negatively correlated with the nodal degree and AUC of nodal efficiency in the left Inferior frontal gurus triangular,MMSE was positively correlated with AUC of nodal efficiency in the left hippocampus.Conclusions:SLE patients have defects in the integrity of the white matter microstructure,and their white matter structural network is also in a suboptimal state.Some abnormal white matter microstructure and structural network indicators are respectively correlated with the scales reflecting disease activity,cognitive function,anxiety,and depression.The above findings suggest that changes in white matter microstructure and its structural network may be one of the pathogenesis of neuropsychiatric symptoms in SLE patients.Part Ⅲ:Study on the Brain Functional Network in Patients with Systemic Lupus ErythematosusObjectives:On the basis of the structural network research of gray matter and white matter,using functional magnetic resonance imaging(fMRI)data,further research on the brain functional connectivity network of SLE based on graph theory and NBS(network based statistic,NBS)research,to explore the network topology characteristics of brain function changes in SLE patients.Methods:DPARFS software was used to preprocess fMRI data,and 116 brain regions were divided according to AAL-116 map as nodes of functional connection network.Firstly,the mean value of BOLD signal intensity in each brain region was calculated,and then the Pearson correlation coefficient between the mean time series in each brain region was calculated.The Pearson correlation coefficient was normalized by fisher-Z transformation and was defined as the connecting edge between each node.Finally,the functional connection matrix of each subject is constructed based on the nodes and edges defined above.Gretna software was used to analyze the functional network based on graph theory,and the differences of global and local attributes of the two groups of brain functional network were compared(FDR correction,q<0.05).NBS software was used for NBS analysis,and independent sample t test was used for inter-group comparison of functional connection network between SLE group and HCs group.The t value was set as 3.5,the number of dislocations was set as 10000,and the significance level was set as 0.01 to find the subnetworks with significant differences between the two groups.GraphPad software was used for Spearman correlation analysis of network indicators with differences between clinical data and intergroup comparison(P<0.05).Results:Compared with the HCs group,Cp,Lp,γ,λ,σ,Eloc was significantly decreased and Eglob was significantly increased in SLE group.In the local attributes,the node degree of SLE group was significantly decreased in the right middle frontal gyrus.There is no correlation between SLEDAI,MMSE,HAMD,HAMA and abnormal global and local attributes.Conclusions:1.Compared with HCs,the whole-brain functional connectivity network of SLE patients showed a significant weakening of global performance and local performance of individual brain regions.In general,like cortical SCNs and white matter structural networks,there was suboptimal network topology properties.2.There is an abnormally weakened sub-network mainly involved in motor management and motor regulation in SLE patients,which provides a reference for revealing the pathogenesis of some neuropsychiatric symptoms including movement disorders in SLE patients.Part Ⅳ:Magnetic resonance based multi-index machine learning in patients with systemic lupus erythematosusObjectives:Using the method of neuroimaging machine learning,a multimodal and multi-index MRI machine learning classification study for SLE was conducted through Support Vector Machines(SVM)to explore the application value and optimal imaging features of machine learning based on brain MRI for SLE diagnosis and classification.Methods:DPARFS software and PANDA software were used to preprocess sMRI,fMRI and DTI data for 50 SLE patients regardless of treatment conditions,27 treatment-naive SLE patients and corresponding HCs to calculate the gray matter volume,Gray matter density,white matter volume,white matter density,ReHo,ALFF,fALFF,and FA images were used as imaging features to perform SVM classification using PRoNTo software,and the leave-one-out cross-validation method was used to verify the classifier,and the weight of contribution of each brain region in the feature to the classification was calculated.127 SLE patients and 102 HCs were classified by NBS-based SVM using NBS-Predict software,the performance of the classification model was evaluated by 10-fold cross-validation and the abnormal sub-networks that could be used for SLE classification under different weight thresholds were calculated.Results:For SLE patients regardless of treatment conditions,the AUC based on the whole brain gray matter density feature classification was 0.97,with an accuracy of 89.00%,and the brain region that contributed the most to the classification was the left paracentral lobule;the AUC based on the whole brain gray matter volume feature classification was 0.97,with an accuracy of 95.00%.The brain region that contributed the most to the classification was the left inferior parietal angular gyrus;the AUC based on the whole brain white matter density feature classification was 0.97,and the accuracy was 87.00%,and the brain region that contributed the most to the classification was the right uncinate tract;based on the whole brain white matter volume feature The AUC of the classification was 0.86,and the accuracy was 89.00%.The brain region that contributed the most to the classification was the right uncinate tract;The AUC based on the ReHo feature classification is 0.71,with an accuracy of 69.00%,and the brain region that contributes the most to the classification is the cerebellar vermis 9;the AUC based on the ALFF feature classification is 0.69,with an accuracy of 65.00%and the brain region that contributes the most to the classification.are the cerebellar vermis 1 and 2;the AUC based on the fALFF feature classification is 0.78,and the accuracy is 67.00%;the AUC based on the FA feature classification is 0.81,and the accuracy is 74.00%,the brain region that contributed the most to the classification was the left superior frontal gyrus,medial orbital;For treatment-naive SLE patients,the AUC based on the whole brain gray matter density feature classification was 1.00,with an accuracy of 94.44%;the AUC based on the whole brain gray matter volume feature classification was 0.99,with an accuracy of 98.18%;the AUC based on the whole brain white matter density feature classification was 0.95,and the accuracy was 92.59%;the AUC based on the whole brain white matter volume feature classification was 0.94,and the accuracy was 88.89%;the AUC based on the ReHo feature classification is 0.86,with an accuracy of 88.33%;the AUC based on the ALFF feature classification is 0.84,with an accuracy of 81.48%;the AUC based on the fALFF feature classification is 0.86,and the accuracy is 83.33%;the AUC based on the FA feature classification is 0.70,and the accuracy is 62.96%.In the classification and recognition of SLE group and HCs group by SVM based on NBS,the AUC of brain functional connection network as feature classification was 0.615,and the accuracy was 0.618.Part 9,left lentiform putamen and other 23 nodes and 41 edge connected sub-networks;when the weight threshold is set to 1,there is a SLE including the right cerebellum 9,cerebellar vermis 9,left lentiform pallidus,etc.12 A sub-network with 12 nodes and 12 edges connected out of tune.Conclusions:1.MRI based imaging machine learning can distinguish SLE patients from healthy controls at the individual level in different clinical application scenarios;2.The performance of the SVM classification model based on sMRI is the best,DTI and fMRI is second,the performance of the SVM classification model based on the functional connection network NBS is weaker,Its significance for SLE classification needs to be supported by more research data;these MRI indicators,especially the brain regions that contribute greatly to the classification,are more meaningful for the classification of SLE,and are potential features for the auxiliary diagnosis of SLE neuroimaging. |
PDF Full Text Request