A Multi-modal Imaging Study On Motor Subtypes Of Parkinson’s Disease Based On Machine Learning Technology

Objective:The purpose of this study was to use support vector machine(SVM)based on machine learning(ML)technology to automate the classification of multimodal MRI data of different motor subtypes of Parkinson’s disease(PD).Mining the neural mechanism under multimodal imaging can promote the differential diagnosis of PD subtypes,provide imaging basis for individualized treatment and optimize clinical diagnosis and treatment methods.Methods:A total of 38 subjects were recruited in this study,including 21 PD patients and17 normal controls(NCs)from the Department of Neurology.All subjects were checked for the relevant scales.The routine MRI sequences,arterial spin labelling(ASL)and Diffusion Tensor Imagings(DTI)sequence scanning were performed using the GE M750 W 3.0T magnetic resonanc scanner.The image was preprocessed and various features were extracted.The classification model was constructed and the classification performance was verified:1.ASL-MRI feature extraction: After applying automated anatomical labelling atlas 3(AAL3)brain region template to ASL image,an image of the same size and shape were obtained.The voxel values of 170 brain regions were obtained according to the location coordinates of brain regions and the data were normalized.2.DTI-MRI feature extraction: The DTI data analysis software(PANDA)based on FMRIB Software Library(FSL)was adopted to scalp the DTI data and set the diffusion matrix parameters.After the diffusion parameters(FA: anisotropy diffusion;MD: average diffusion rate;AD: axial diffusion index;RD: radial diffusion index)calculated and the standard template registered,we resample and smooth the diffusion parameters.Finally,we select the white matter partitioning map template and calculate the diffusion index based on the white matter map.The AAL3 template was used for registration with the white matter atlas image of DTI to obtain the same size and shape image.The voxel values of 170 brain regions were obtained according to the location coordinates of brain regions and the data were normalized.3.ASL-DTI fusion feature extraction: The voxels of each brain region of ASL were fused with the voxels of four feature parameters of DTI(FA,MD,AD and RD),and the voxel data of different features in the same brain region of the same subject were normalized.4.Model classification and validation: according to the obtained voxel values of brain regions,the patients were classified in each brain region.Based on the data of NC,PIGD patients and TD patients,three binary classification models are proposed: "NC vs.Others","PIGD vs.Others",and "TD vs.Others".SVM classifiers of three binary models were constructed and the performance of the classifiers was estimated using the Leave-one-out cross-validation(LOOCV)method.The classification performance of SVM model was evaluated by model performance indexes of accuracy,sensitivity,specificity and maximum area under curve(AUC)measured in ROC curve analysis.Results:1.Classification performance evaluation of ASL-MRI: After SVM screening,NC and PD could be distinguished well by the subgenual of anterior cingulate cortex,and a high diagnostic rate was obtained(accuracy =92.31%,AUCmax =0.9585).The TD group had the highest diagnostic accuracy in the right supramarginal gyrus(accuracy =84.21%,AUCmax =0.9192).The diagnostic accuracy of the right intralaminar of thalamus in distinguishing PIGD from other groups was the highest(accuracy =89.47%,AUCmax =0.9464).2.Classification performance evaluation of DTI-MRI: In the “NC vs.Others”model,the characteristic brain regions involved thalamus-related brain regions,putamen,right nucleus accumbens,right locus coeruleus,limbic gyrus and right supplementary motor area.The AD of the right nucleus accumbens and the RD of the right lateral geniculate of thalamus were the best classification features with accuracy of 84.21%.Thalamic-related brain regions and the lobule VI of vermis were the characteristic brain regions of PIGD.The RD of the left intralaminar of thalamus was86.84% accurate and the AUC value was 82.14%,which was the best characteristic to distinguish PIGD from other groups.PIGD was correlated with the MD,AD and RD values of the lobule VI of vermis,which could be distinguished from other groups to a certain extent.The left Crus I of cerebellar hemisphere,locus coeruleus,median raphe nuclei and thalamic related brain areas were the characteristic brain areas of TD.The AD value of the right locus coeruleus was the best distinguishing feature with 78.95%accuracy and 77.78% AUC value.3.Classification performance evaluation of ASL-DTI fusion: The right nucleus accumbens was the best classification feature with 81.58% accuracy and 87.68%AUC value.The left intralaminar of thalamus,right pulvinar medial of thalamus and right parahippocampal gyrus became the characteristic brain regions of PIGD.The accuracy of the left thalamic lamina nucleus reached 81.58%,which could be distinguished from other groups to a certain extent.The left rectus gyrus and the right anterior ventral thalamic nucleus were the characteristic brain regions of TD.The left rectus gyrus was the best feature to distinguish TD from other groups with an accuracy of 78.95% and an AUC value of 75.08%.Conclusions:1.SVM based on ML technology can be used to automatically classify multimodal image data of different motor subtypes of PD.2.The classification performance of the right supramarginal gyrus and the right intralaminar of thalamus was consistent with the blood perfusion changes of the subtypes,and the blood perfusion patterns of the PD subtypes may be different.3.The right nucleus accumbens,limbic gyrus,right locus ceruleus,putamen and right supplementary motor area are expected to be neuroimaging markers for PD and NC classification.The left intralaminar of thalamus can be the best characterized brain region for PIGD.The left Crus I of cerebellar hemisphere and the right locus coeruleus are helpful in investigating the structural patterns of TD circuits.4.With the fusion of ASL and DTI,the right nucleus accumbens became a stable neuroimaging marker to distinguish PD from NC,and the left intralaminar of thalamus and the left rectus gyrus were the best features of PIGD and TD respectively.