Support Vector Machine Broad Learning System For Early Diagnosis Of NPSLE By Multimodel MRI

Objective: Magnetic resonance spectroscopy(MRS),diffusion kurtosis imaging(DKI),and voxel-based morphometry(VBM)based on the theory and method of machine learning were evaluated.By comparing the diagnostic efficiency of different machine learning model with different models of magnetic resonance imaging,we aim to find the best diagnosis model for the early diagnosis of neuropsychiatric lupus erythematosus(NPSLE).Method: All subjects were divided into the NPSLE group and the healthy control group.Both of the two groups underwent a 3.0T magnetic resonance imaging with routine MRI,MRS,DKI,and three-dimensional brain volume imaging(3D-BRAVO)sequences.The data were analyzed to obtain the absolute concentration and their ratio of N-acetylaspartate(NAA),N-acetylaspartate glutamate(NAAG),choline(GPC / Cho),phosphocholine(PCr),creatine(Cr),phosphocreatine(PCr),glutamate(Glu),glutamine(Gln)and myo-inositol(m I/Ins)in bilateral posterior cingulate gyrus(PCG),dorsal thalamus(DT),lentiform nucleus(LN),posterior horn of the lateral ventricle paratrigonal white matter(PWM)and right insular(RI).Furthermore,we calculated the value of fractional anisotropy(FA),mean diffusion tensor(MD),and mean kurtosis(MK)in bilateral PCG,DT,LN,and PWM,from the raw data of DKI.We also measured the volume value of the gray matter,white matter,and cerebrospinal fluid using VBM.Next,to propose a machine learning model.(1)The model was used to combine with MRS,DKI,or VBM respectively,to establish a single-model machine learning model for distinguishing NPSLE from HC.(2)There were three dual-model machine learning models be established,including(MRS+DKI)–(BL-SVM),(MRS+VBM)-(BL-SVM)and(DKI+VBM)-(BL-SVM).(3)A multi-model machine learning model((MRS+DKI+VBM)-(BL-SVM))also was established to diagnose NPSLE.Results: The support vector machine based on broad learning system(BL-SVM)was obtained in this study.The MRS single-model machine learning model established by BL-SVM distinguishes healthy controls from neuropsychiatric lupus erythematosus with an accuracy,sensitivity,and specificity of 95%,100% and 90%,respectively.The accuracy,sensitivity and specificity of DKI single-mode machine learning model were 59%,78% and,40%,respectively.The accuracy,sensitivity and specificity of VBM single-mode machine learning model were58%,78% and,45%,respectively.The accuracy,sensitivity and specificity of(MRS+DKI)–(BL-SVM)were 95%,100% and 90% respectively.The accuracy,sensitivity and specificity of(MRS+VBM)–(BL-SVM)were 85%,95% and 75% respectively.The accuracy,sensitivity and specificity of(DKI+VBM)–(BL-SVM)were 68%,81.7% and 55% respectively.The multimodal magnetic resonance machine learning model,which combines the multimodal magnetic resonance data of MRS,DKI and VBM as input features,was used to distinguish healthy controls from neuropsychiatric lupus erythematosus.The accuracy,sensitivity and specificity were 97.5%,100% and,95%,respectively.Conclusions: The BL-SVM based on multimodal magnetic resonance data has the best diagnostic efficiency.However,the main contribution comes from the characteristics represented by MRS,and the characteristics represented by DKI and VBM are too low in sensitivity and specificity for the diagnosis of NPSLE.Therefore,MRS combined with a support vector machine based on broad learning system is the best NPSLE diagnosis model obtained in this study.It provides a non-invasive,non-radiation,a convenient,and objective diagnostic method for the early diagnosis of NPSLE in the future.