Differential Diagnosis of Parkinsonism with Deep Metabolic Imaging Biomarker – An Artificial Intelligence-Aided Multi-Cohort FDG PET Study
Ping Wu,Yu Zhao,Jianjun Wu,Jingjie Ge,Ling Li,Jiaying Lu,Ian Alberts,Igor Yakushev,Qian Xu,Yimin Sun,Fengtao Liu,Matthias Brendel,Günter U. Höglinger,Claudio Bassetti,Yihui Guan,Wolfgang H. Oertel,Wolfgang Weber,Axel Rominger,Jian Wang,Chuantao Zuo,Kuangyu Shi
DOI: https://doi.org/10.2139/ssrn.3675448
2020-01-01
SSRN Electronic Journal
Abstract:Background: The clinical presentation of early idiopathic Parkinson’s disease (IPD) substantially overlaps with that of atypical parkinsonian syndromes such as multiple system atrophy (MSA) and progressive supranuclear palsy (PSP). 18F-fluorodeoxyglucose (FDG) positron emission tomography (PET) has proven useful in assisting in the differential diagnosis of these parkinsonian syndromes. The aim of this study was to develop a metabolic imaging biomarker based on deep learning to support the differential diagnosis of parkinsonism. Methods: In addition to 863 non-parkinsonian subjects, a total of 1275 parkinsonian patients were accessed who underwent FDG PET imaging, clinical evaluation, and follow-up, yielding clinically possible, definite, or confirmative diagnoses of IPD, MSA, and PSP. A 3D deep convolutional neural network was developed to extract a deep metabolic imaging (DMI) biomarker. The accuracy of the DMI biomarker was evaluated with a cross-validation and a blind test. The robustness of the DMI biomarker was assessed on patients with longitudinal follow-up. The functional basis of DMI biomarker was further investigated based on the saliency map of the obtained deep neural network. Findings: The proposed DMI biomarker achieved an area under the receiver operating characteristic curve of 0·986 for the differential diagnosis of IPD, 0·997 for MSA, and 0·982 for PSP in the cross-validation. In the blind test, the DMI biomarker resulted in sensitivities of 98·1%, 88·5%, and 84·5%, as well as specificities of 90·0%, 99·2%, and 97·8% for IPD, MSA, and PSP respectively. The integration of demographic information and clinical assessments to DMI biomarker resulted in slight improvements. Saliencies for the DMI biomarker were found in several parkinsonism-related regions such as midbrain, putamen, and cerebellum. Interpretation: The DMI biomarker based on deep learning shows potential to provide accurate differential diagnosis for parkinsonism on FDG PET. The functional regions behind the DMI biomarker are consistent with known parkinsonian pathology.Funding Statement: The work was supported by grants from the National Natural Science Foundation of China (No. 81771483, 81671239, 81361120393, 81401135, 81971641, 81902282, 91949118, 81771372), from the Ministry of Science and Technology of China (2016YFC1306504), Shanghai Municipal Science and Technology Major Project (No. 2017SHZDZX01, 2018SHZDZX03) and ZJ Lab, Shanghai Sailing Program (No. 18YF1403100). It was also supported by Swiss National Science Foundation (No. 188350) and Siemens Healthineers.Declaration of Interests: W.H.O is Hertie Senior Research Professor, supported by the Charitable Hertie Foundation, Frankfurt/Main, Germany. A.R. and K.S. received research support from Siemens Healthineer. Other authors report no financial interests or potential conflicts of interest regarding the subject matter of the manuscript.Ethics Approval Statement: All procedures performed in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. All the data were from Parkinson's Disease Database and Samples Bank in Huashan Hospital. Ethics permission was obtained from the Institutional Review Board of Huashan Hospital and written consent was obtained from each subject after detailed explanation of the procedures.