Machine Learning Based Multi-modal Magnetic Resonance Imaging Studies In Patients With Idiopathic Parkinson’s Disease And Parkinsonian Variant Of Multiple System Atrophy

Objective: Idiopathic Parkinson’s disease(IPD)and multiple system atrophy(MSA)are two general neurodegenerative diseases.According to clinical symptoms,IPD can be classified into tremor-dominant(TD)subtype and postural instability and gait disability(PIGD)subtype.MSA can be divided into parkinsonism subtype(MSA-P)and cerebellar subtype(MSA-C).IPD and MSA,especially MSA-P,share similar clinical symptoms,causing high misdiagnosis rate.However,the disease progression,treatment and prognosis are different between two diseases,thus an early differential diagnosis is of great importance.In addition,IPD is considered as a highly heterogeneous disease with the development of research.Therefore,refinement in the clinical classification of IPD subtype may contribute to individualized treatment.Multimodal magnetic resonance imaging(m MRI)utilizes advantages and complementarities of sequences sensitive to different tissue in order to obtain complete characterization of comprehensive alteration of disease,comprising three dimensional T1 weighted imaging(3DT1),blood oxygen level dependent functional MRI(BOLD f MRI),susceptibility weighted imaging(SWI),diffusion tensor imaging(DTI).Traditional m MRI analysis is based on group-level method mostly.Inconsistencies have been reported due to differences in scan parameters and postprocessing approaches,rendering it hard to clinical transformation.As a result,how to perform individualized prediction for new individuals based on m MRI imaging biomarkers requires to be solved.Besides,how to extract valuable information from high dimensional medical imaging and to make combination analysis of multimodal information is challenging.The combination of artificial intelligence and medical imaging solved above questions.To date,clinical applicable automatic differential diagnosis model between IPD and MSA-P and IPD subtyping model is scarce.Therefore,this study is to build differential diagnosis of IPD and MSA-P model and IPD subtyping model based on m MRI data,in order to achieve individualized diagnosis and treatment.Meanwhile,to find the differential diagnosis biomarkers to unravel distinct brain structural and functional alterations between IPD and MSA-P and between different motor subtypes of IPD.Participants and Methods: 96 IPD and 102 MSA-P patients were included in the study.The clinical diagnosis of IPD was based on the UK PD Society Brain Bank diagnostic criteria and MSA was according to the second consensus clinical criteria.According to UPDRS score,IPD patients were divided into TD and PIGD subtype.All patients underwent 3.0 T MR scanner(Siemens,Verio)to obtain m MRI data,including SWI,BOLD f MRI,3DT1,DTI sequences.Meanwhile,all patients underwent clinical assessments,including UPDRS,Hoehn-Yahr,MMSE,Mo CA measurements.1.SWI-based radiomics of basal nuclei in differentiating IPD from MSA-P:Extracting radiomic features from established basal nuclei.The most informative feature subsets were defined after feature selection methods.Then differential diagnosis models were built.Furthermore,clinical measurement was added into the model.2.Striatum-based multimodal MRI machine learning model in differentiating IPD from MSA-P: Extracting volume,functional and DTI measures from striatum subdivisions.The most informative feature subsets were defined after feature selection methods.Then differential diagnosis models were built.Furthermore,SHAP analysis was applied to interpret model prediction.3.Multilevel indices of rs-f MRI machine learning model in IPD subtyping: Extracting multilevel indices of rs-f MRI from whole brain regions.The most informative feature subsets were defined after feature selection methods.Then differential diagnosis models were built.Furthermore,SHAP analysis was applied to interpret model prediction.Results: 1.SWI-based radiomics of basal nuclei in differentiating IPD from MSA-P: The radiomic model based on putamen showed the best performance in differential diagnosis of IPD and MSA-P,with accuracy of 0.810 and AUC of 0.867 in the training dataset,and accuracy of 0.791 and AUC of 0.862 in the validation dataset.Furthermore,a combined model,which incorporated UPDRSIII measurements into the radiomic model,further improved performance,with accuracy of 0.836 and AUC of 0.880 in the training dataset and accuracy of 0.809 and AUC of 0.878 in the validation dataset.2.Striatum-based multimodal MRI machine learning model in differentiating IPD from MSA-P: Compared with single modality model,multimodal model based on striatal volume,functional measures and DTI scalars manifested the best performance in differential diagnosis of IPD and MSA-P,with accuracy of 0.897 and AUC of 0.937 in the training dataset and accuracy of 0.820 and AUC of 0.899 in the validation dataset.SHAP analysis showed that functional activity of left dorsolateral putamen was the most valuable feature in model prediction.Besides,DTI scalar of dorsolateral putamen and the functional connectivity between dorsolateral putamen and frontal lobe were also important for prediction.3.Multilevel indices of rs-f MRI machine learning model in classification of motor subtype of IPD: Compared with separate index model,multilevel indices of rs-f MRI model showed optimal performance,with accuracy of 0.932 and AUC of 0.934 in the training dataset and accuracy of 0.875 and AUC of 0.917 in the validation dataset.SHAP analysis showed that functional activity and connectivity of frontal lobe and cerebellum played important role in model prediction.Conclusions: Our study established automatic differential diagnosis and subtyping models based on the combination of machine learning and multimodal MRI.The results showed that diagnosis model based on multilevel and multimodal MRI had strong robustness and generalization,rendering high clinical application potential.Besides,the analysis of model prediction showed that IPD and MSA-P patients had different putamen iron deposition and striatum functional and structural alterations patterns.Compared with IPD,MSA-P had heterogeneous iron deposition and more severe functional and structural damage in putamen.The dysfunction of putamen further caused frontal dysfunction via striatum-cortical pathway,which might provide explanation for faster progression and poorer prognosis in MSA-P patients.In addition,different IPD subtypes had distinct functional alterations in frontal lobe and cerebellum,which might be one of the reasons behind the more severe clinical manifestations and poorer prognosis in PIGD subtype.