Use of magnetic resonance imaging and artificial intelligence in

2 days ago - Parkinson's disease (PD) is a common neurodegenerative disorder. It has a delitescent onset and a slow process. The clinical manifestatio...
0 downloads 0 Views 1015KB Size
Subscriber access provided by Bibliothèque de l'Université Paris-Sud

Review

Use of magnetic resonance imaging and artificial intelligence in studies of diagnosis of Parkinson's disease Jingjing Xu, and Minming Zhang ACS Chem. Neurosci., Just Accepted Manuscript • DOI: 10.1021/ acschemneuro.9b00207 • Publication Date (Web): 14 May 2019 Downloaded from http://pubs.acs.org on May 14, 2019

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 23 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Chemical Neuroscience

Use of magnetic resonance imaging and artificial intelligence in studies of diagnosis of Parkinson's disease Authors: Jingjing Xu, and Minming Zhang* Corresponding to Minming Zhang, Department of Radiology, the Second Affiliated Hospital of Zhejiang University, School of Medicine. No.88 Jiefang Road, Shangcheng District, Hangzhou, China, 31000. E-mail: [email protected]

Abstract Parkinson’s disease (PD) is a common neurodegenerative disorder. It has a delitescent onset and a slow process. The clinical manifestations of PD patients are highly heterogeneous. Thus PD diagnosis process is complex which mainly depends on the professional knowledge and experience of the physician. Magnetic resonance imaging (MRI) could detect the small changes in the brain of PD patients and quantitative analysis of brain MRI may improve the clinical diagnosis efficiency. However, due to the complexity of clinical courses in PD and the high dimensionality in multimodal MRI data, traditional mathematical analysis could not effectively extract the huge information in them. Up to now, the accuracy of PD diagnosis in large sample size is still unsatisfying. As the artificial intelligence (AI) is becoming more mature, varieties of statistical models and machine learning (ML) algorithms were used for quantitative imaging data analysis to explore a diagnosis result. This review aims to state an overview of existing research recently that used statistical ML/AI methods to perform quantitative analysis of MR image data for the study of PD diagnosis. First we review the recent research in three sub-areas: diagnosis, differential diagnosis and subtyping of PD. Then we described the overall workflow from MR image to classification result. Finally, we summarized a critical assessment of the current research and provide some recommendations for likely future research developments and trends.

Keywords Magnetic resonance imaging, Artificial intelligence, Parkinson’s disease, Machine learning

ACS Paragon Plus Environment

ACS Chemical Neuroscience 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Introduction Parkinson’s disease (PD) is the second most common neurodegenerative disorder in the world [1]. The prevalence of PD in Chinese elderly people is about 1.7% [2], and over 3 million Chinese presently suffer from it. Worldwide, an estimated 15 million people are thought to be affected [3]. The impact of PD on elderly patients’ health is enormous, and growing with the development of social aging, which makes finding accurate and early diagnosis method to reduce burdens of this disease an extremely important priority in China and worldwide. Currently, the clinical diagnosis of PD depends mainly on the diagnostic criteria of the UK Parkinson's Disease Society Brain Bank (PDSBB) or International Parkinson and Movement Disorder Society (MDS) [4, 5]. In PDSBB/MDS-PD criteria, the motor syndrome remains the core feature by which clinical PD is defined. Many patients cannot be diagnosed prior to the classical motor phase of PD [6]. Meanwhile, the clinical manifestations of PD patients are highly heterogeneous. The patient can be divided into three subtypes according to their clinical symptoms: tremor dominant (TD), bradykinesia/Akineto-rigid (AR) and postural instability and gait difficulty (PIGD) [7-9]. Different phenotypical subtypes have varying responses to treatment and different prognoses [10, 11]. In clinical practice, it is also necessary to differentiate PD from multiple system atrophy (MSA) and progressive supranuclear palsy (PSP). They are clinically similar yet different entities [12, 13]. So far, PD diagnosis process is complex. It depends too much on the expertise in defining patients’ symptoms, especially in prodromal stage when PD patients may present with a variety of non-motor symptoms and/or subtle motor signs that do not meet current diagnostic criteria. Further, the current diagnostic criteria are not effective in classifying the subtypes of the disease and differential diagnosis of PD. According to some systematic reviews, the clinical diagnostic accuracy was not satisfying (about 80%) and thus PD diagnosis is still challenging [14]. Therefore, it is of great significance to find objective quantitative indicators to promote clinical diagnosis. During the past decade, the use of imaging as a strategy to diagnosis PD has been significantly highlighted. Imaging biomarkers are imminently required to improve the accuracy of clinical diagnosis PD and to assess early disease in vivo [14, 15]. The most mature imaging markers of PD have targeted the hallmark pathology in the substantia nigra dopaminergic system [16, 17]. As a result, positron emission tomography (PET) and single photon emission computed tomography (SPECT) measures are very effective in the motor phase. But it would be desirable to diagnose the disease in the early stages, ideally before the degeneration of dopaminergic neurons in the substantia nigra. With this regard, it has been recognized that structural magnetic resonance imaging (sMRI) as well as advanced MRI techniques, including resting-state functional MRI (rs-fMRI), diffusion tensor (DTI), and magnetic sensitive imaging (SWI), have enabled to identify subtle neuronal injury in neurodegenerative diseases [18, 19]. In PD researches, voxel based morphometry (VBM) has revealed progressive atrophy in the early stage of the disease [20, 21]. Disrupted functional connectivity and microstructural changes have been detected by rs-fMRI and DTI [22, 23]. These different modalities of MRI play different but complementary roles. Multimodal MRI can provide structural, functional and metabolic information of the brain change. It remains unknown, however, how to define valid MRI markers which diagnose or evaluate the disease before the damage of dopaminergic neurons.

ACS Paragon Plus Environment

Page 2 of 23

Page 3 of 23 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Chemical Neuroscience

While the use of multimodal MRI for diagnosis PD is primarily based on subtle structural and functional abnormalities of the brain in PD patients, researchers have developed and are exploiting advanced statistical models and machine learning (ML) algorithms for quantitative imaging data analysis to make a classification result for diagnosis of PD. In recent years, data mining, neural network, deep learning and other mathematical methods are developing at an unprecedented speed. They have been widely used in the field of images analysis, and showed great potential in analysis of medical images. The application of these new methods may further improve the analysis ability of complex multimodal MR image data and improve the efficiency of PD diagnosis. It may be a promising tool to assist in automatic diagnosis, integrated evaluation and advancing mechanism research of neurological and psychiatric diseases. This article aims to review existing studies on quantitative analysis of different modal MR image data using statistical/ML/AI methods in diagnosis of PD. We used PubMed as the search engine. Only articles published in English were considered for the review. We try to review the existing work focusing on diagnosis, differential diagnosis and subtyping; to provide an overall workflow of AI, to analyze the affected brain regions and to conclude the article and propose some future research directions.

Use of different MRI modalities in studying PD diagnosis The core pathological change of PD is degeneration and death of dopaminergic neurons in substantia nigra [24, 25], which leads to degeneration of basal ganglia nuclei [16]. Our team applied different modalities of MRI including sMRI, fMRI, iron imaging to detect the neurodegenerative abnormality of PD patients [22, 26-34]. Among existing MR modalities, brain iron imaging and neuromelanin imaging were recently preferred in demonstrating the degeneration of substantia nigra. Iron imaging showed that the dorsolateral hyperintense of substantia nigra disappeared in PD patients [35-37], indicating the abnormal deposition of brain iron in the nuclei. Neuromelanin imaging in PD patients showed that the volume and signal intensity of locus coeruleus and substantia nigra compact reduced [38]. Different MRI modalities can detect the structural and functional abnormalities of the brain in PD patients from multiple perspectives and some may reflect neurodegenerative changes to a certain extent. However, how to integrate these MR research results to assist PD clinical diagnosis is yet to known. Diagnosis of PD using machine learning methods based classification algorithms is then referred and we briefly reviewed the studies as follows with a focus on the role of different MRI modalities and ML methods. A classical approach adopted by earlier studies for diagnosis PD was to compute features from single modality or multiple modalities into a feature set, and a classifier was built then based on this set. Features were usually selected based on group differences. For example, Tang et al. used the amplitude of low-frequency fluctuation (ALFF) and the fractional ALFF (fALFF) as features and used support vector machine (SVM) and the leave-one-out cross-validation method to build a 51 PD vs. 50 normal controls (NC) classifier, which had an 84.2% correct classification rate [39]. Long et al. made a feature set that composed of features from resting-state functional magnetic resonance

ACS Paragon Plus Environment

ACS Chemical Neuroscience 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

imaging (rsfMRI) and structural images. A SVM classifier was built after feature selection. Their approach achieved 86.96% accuracy in classifying 19 early Parkinson's disease and 27 normal a volunteers based on leave-one-out cross-validation method [40]. New and robust methods to extract features related to a disease and to remove noise features were later developed. Chen et al. used scale invariant feature transform (SIFT) algorithm to extract features from the brain sMRI and developed a novel local feature based SVM approach to detect brain changes in PD. The classifier achieved 80% accuracy in 9 patients and 6 healthy controls from the PPMI dataset [41]. In another work, Chen et al. considered the whole-brain resting-state functional connectivity patterns. The Kendall tau rank correlation coefficient was performed to extract the final feature set for classification. They used a SVM with linear kernel function to construct the classifier, achieved 93.62% accuracy in classifying 21 PD patients with 26 demographically matched healthy controls based on leave-one-out cross-validation [42]. We speculate that the analysis of disease-affected brain regions provides us a consistent way of diagnose of PD. Also proper approaches of selecting features and combination of multidimensional features from different MR modalities should provide a more reasonable approach which may further improve PD diagnosis accuracy. Later, researchers attempted to select the most discriminative features, along with the larger samples to build a classifier. Adeli et al. used a joint feature-sample selection (JFSS) removing poor samples and irrelevant features to build a linear regression model that can classify 374 PD and 169 NC. Volumetric features were extracted and selected from 98 labeled ROIs. The linear discriminant analysis (LDA) classifier achieved a classification accuracy of 81.9% on the PPMI database using 10-fold cross validation [43]. In their following study, a joint kernel-based feature selection was applied to pick a subset of discriminative sMRI features to best construct a model. The model can classify 369 PD and 169 NC with a classification accuracy of 70.5% [44]. In the study of Pläschke et al. [45], the SVM classification of 80 PD patients and 84 matched healthy controls was performed on functional connectivity in 12 task-based, meta-analytically defined networks using 25 replications of a nested 10-fold cross-validation scheme. Rs-fMRI networks subserving autobiographical or semantic memory, motor execution, and theory-of-mind cognition achieved the best classification of PD with accuracy 67%-75%. Large datasets created a large number of features, so researchers developed novel frameworks to analyze substantial image features. Peng et al. combined low-level ROI features (gray matter volume, cortical thickness, etc.) and high-level correlative features (connectivity between ROIs) to frame the multilevel ROI features to explore sensitive morphometric biomarkers, and applied the filter- and wrapper- based features extraction approaches and the multi-kernel SVM for classification between 69 PD patients and 103 normal controls from PPMI dataset [46]. Their model achieved an accuracy of 85.78%, a specificity of 87.79%, and a sensitivity of 87.64%. Then, imaging and non-imaging data was aimed to aggregate and harmonize for model building. In the study by Dinov et al. [47], PPMI neuroimaging, clinical, demographic and genetic data was concatenate to develop classification and predictive models for Parkinson’s disease. They employed model-based and model-free methods for predictive analysis, and the accuracy of model-free machine learning based classification in large-scale subjects was over 80%. Amoroso et al. used Random Forests (RF) for feature selection and a SVM to integrate network features with typical clinical data of PD prodromal phase and provide a diagnostic classifier of 374 PD and 169 NC subjects from PPMI dataset. The combined use of network features (NF) and clinical features (CF) achieved 93% accuracy, 93% sensitivity and 92% specificity [48]. Ideally,

ACS Paragon Plus Environment

Page 4 of 23

Page 5 of 23 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Chemical Neuroscience

a classification analysis should integrate multiple MR modalities with multi-level features e.g., T1w VBM, fMRI, DTI and SWI, preferably combined with other parameters such as demographics, gait analysis, neuropsychological tests, laboratory tests, etc., with the purpose to further increase the classification accuracy and robustness of individual diagnosis. Some studies focused on comparing different feature selection methods in the Parkinson’s biomarker identification. For example, Liu et al. introduces folded concave penalized (FCP) sparse logistic regression to identify 15 multimodality MRI markers and the University of Pennsylvania Smell Identification Test (UPSIT) as biomarkers for PD diagnosis from a great quantity of potential factors [49]. They compared FCP with the basic, the Lasso penalized scheme and the principle component analysis (PCA)-based feature selection in the Parkinson's biomarker identification. Among the four approaches, the FCP-based approach was the best performer (99.7% accuracy in classifying 34 PD and 31 NC), while the PCA-based approach had the lowest accuracy (77.8%). Most existing work as reviewed previously classifying PD patients from NC highlighted the important roles of the basal ganglia and cerebral motor cortex, the critical regions known to be strongly involved in the pathophysiological mechanisms of PD. And the previous discriminative features were clinically and neurobiologically related to these target brain regions. A brief summary can be found in Table 1. Although ML studies using rsfMRI revealed cerebellum functional alteration involvement in the diagnosis of PD [39, 42], the potential of cerebellar morphological abnormalities have not been tested in the diagnosis of PD. Recently, Zeng et al. [50] investigated the gray matter changes in the cerebellum in PD patients. They used SVM with voxel-based morphometric and classified the possible PD patients from NC with accuracy of more than 95 % via cross-validation. The results suggested the potential of the cerebellar structural alterations in PD patients, which may assist in PD diagnosis. It is essential to systemically investigate these structural and functional alterations of the cerebellum in probable PD patients in the future. A larger sample size and multicenter imaging data should be considered. Table 1 Summery of studies using magnetic resonance imaging and artificial intelligence in studies of diagnosis of Parkinson's disease Papers

MRI Modali ties

Data source

Subjects

Feature extractio n method

Validati on method

Classifie r

Acc.

Affected brain regions

Tang et al. 2017 [39]

rsfMRI

recruit ed

51 PD/ 50 NC

group differenc e analysis

LOOCV

SVM

84.2%

the lingual gyrus, putamen, cerebellum posterior lobe

Long et al. 2012 [40]

Multimodalit y: rsfMRI, sMRI

recruit ed

19 PD/ 27 NC

group differenc e analysis

LOOCV

SVM

86.96 %

ORBmid, ROL, PHG, ANG, MTG, PCL, PreCG, PCG

Chen et al. 2014

sMRI

PPMI

9 PD/ 6 NC

scale invariant

LOOCV

SVM

80%

Limbic lobe, Frontal lobe, Sub-lobar,

ACS Paragon Plus Environment

ACS Chemical Neuroscience 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

[41]

Page 6 of 23

feature transfor m

Midbrain, Posterior Occipital lobe

Pons, lobe,

Chen et al. 2015 [42]

rsfMRI

recruit ed

21 PD/ 26 NC

Kendall tau rank correlati on coefficie nt

LOOCV

SVM

93.62 %

DMN, the control network, the cerebellum, etc

Adeli et al. 2016 [43]

sMRI

PPMI

374 PD /169 NC

joint featuresample selection

10-fold crossvalidatio n

LDA

81.9%

the red nucleus, substantial nigra, pons, middle frontal gyrus, superior temporal gyrus

Adeli et al. 2017 [44]

sMRI

PPMI

369 PD /169 NC

joint kernelbased feature selection

10-fold crossvalidatio n

SVM

70.5%

insula, cingulate gyrus, hippocampus, parahippocampal gyrus, amygdala, caudate, putamen, etc

Pläschk e et al. 2017 [45]

rsfMRI

recruit ed

80 PD /84 NC

previous quantitat ive metaanalyses

10-fold crossvalidatio n

SVM

75%

rsfMRI networks subserving autobiographical or semantic memory, motor execution, and theory-of-mind cognition

Peng et al. 2017 [46]

sMRI

PPMI

69 PD/103 NC

filterand wrapperbased features extractio n

10-fold crossvalidatio n

SVM

85.78 %

frontal lobe, parental lobe, limbic lobe, temporal lobe, and central region

Dinov et al. 2016 [47]

sMRI

PPMI

263 PD/123 NC

the Rpackage CARET

5-fold crossvalidatio n

SVM, AdaBoo st

over 80%

superior parietal gyrus, putamen, caudate

Amoros o et al. 2018 [48]

sMRI

PPMI

374 PD/169 NC

random forests

10-fold crossvalidatio n

SVM

93%

frontal, occipital and temporal lobes, Limbic Lobe, brainstem. midbrain

Liu et al. 2016 [49]

Multimodalit y:

recruit ed

34 PD/31 NC

FCP, Lasso, PCA

bootstra pping

FCP

99.7%

amygdala, caudate, substantia nigra, pallidus, hippocampus,

ACS Paragon Plus Environment

Page 7 of 23 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Chemical Neuroscience

sMRI, R2*, DTI Zeng et al. 2017 [50]

sMRI

red nucleus, dentate

recruit ed

45 PD/40 NC

recursive feature eliminati on

crossvalidatio n

SVM

>95%

the cerebellar Crus I

Acc.: Accuracy, SVM: support vector machine, PD: Parkinson's disease, NC: normal controls, LOOCV: leave-one-out cross-validation method, rsfMRI: resting state functional MRI, sMRI: structural MRI, ORBmid: Middle frontal gyrus, orbital part, ROL: Rolandic operculum, PHG: Parahippocampal gyrus, ANG: Angular gyrus, MTG: Middle temporal gyrus, PCL: Paracentral lobule, PreCG: Precentral gyrus, PCG: Posterior cingulate gyrus, DMN: the default mode network, LDA: linear discriminant analysis, FCP: folded concave penalized, PCA: principle component analysis-based feature selection.

Use of different MRI modalities in studying PD differential diagnosis The most urgent and difficult clinical problem is not the classification of PD patients from healthy controls but the discrimination of idiopathic PD versus other atypical diseases of Parkinsonism such as MSA and PSP. MRI studies show that there are different brain imaging changes in these diseases. For example, MSA patients may present with “Hot cross bun” sign [51], decreased volume of cerebellum, thalamus, putamen and brainstem [52, 53], diffusion pattern change [54, 55], and altered functional and causal connectivity of cerebello-cortical circuits [56]. PSP patients may present with “hummingbird” sign, “Mickey Mouse” sign [57], and diffusion pattern changes in the midbrain and basal ganglia [58, 59]. In addition, significant iron deposition was found in the posterior putamen and occipital thalamus in MSA patients [60-62]. These MRI features extracted from T1WI, DTI and SWI may be helpful for the differential diagnosis of PD. Researchers have tried different MR features to train a classification model. sMRI was the most used modality. Salvatore et al. [63] developed classifier for individual diagnosis of 28 PD versus 28 PSP by sMRI. They used a supervised machine learning algorithm to obtain voxel-based morphological biomarkers. The algorithm was based on the combination of PCA as feature extraction technique and on SVM as classification algorithm. The model obtained an accuracy, specificity and sensitivity>90% in classifying PD from PSP. sMRI features were also used by Focke and his colleagues [64]. They performed a SVM classification with leave-one-out cross-validation for an individual diagnosis in PSP versus PD with up to 96.8% accuracy. Meanwhile in MSA versus PD classification, the algorithm achieved an accuracy of 71.9%. MRI volumetric feature analysis combined with SVM classification was further considered in a multicenter study with a large cohort of subjects (204 PD, 106 PSP, 81 MSA and 73 NC) [65]. Differentiation of PD versus other Parkinsonism was not bad (>80%). Besides sMRI, other MR modalities have been already applied in ML/AI-based studying PD differential diagnosis. In the study by Haller et al. [66], increased SWI signal intensity was demonstrated in the bilateral thalamus and left substantia nigra in PD patients.

ACS Paragon Plus Environment

ACS Chemical Neuroscience 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

At the individual level, SVM correctly classified PD patients versus other Parkinsonism with an accuracy above 86 %. Haller et al. studied the tract-based spatial statistics (TBSS) preprocessed DTI fractional anisotropy (FA) data in a cohort of 17 PD and 23 other Parkinsonism. They selected the top hundreds of features with 10-fold cross-validation. SVM analysis of FA provided a classification between PD versus other Parkinsonism with accuracies of up to 97% [67]. Cherubini et al. [68] used DTI and sMRI VBM in a SVM algorithm to differential diagnose 21 PSP and 57 PD. DTI parameters combined with gray matter atrophy features demonstrated 100% accuracy, 90% sensitivity and 96% specificity. They believed that the microstructural tissue alterations exhibited by DTI were somehow complementary to gray matter (GM) atrophy measured by VBM for distinguishing PSP from PD. According to these studies, classification between PD and PSP patients showed good effectiveness. Critical regions (corpus callosum, midbrain, pons, and thalamus) known to be affected in the pathophysiological mechanisms of PSP were mostly involved in feature analysis and classification. sMRI was still the mainstream of MR modalities selected in those studies. But feature extraction and selection methods were varied. A majority of existing work as reviewed above focused on using different MRI modalities to distinguish PD patients from other Parkinsonism. A few researchers explored the potential usefulness of machine learning for differentiating PD patients with special symptoms or certain phase. Cherubini et al. [69] used a multi-kernel approach to build a classifier, which differentiated between patients who have essential tremor (ET) with resting tremor (rET) from those who have tremor-dominant Parkinson’s disease (tPD). For MRI processing, a combined whole-brain, VBM and DTI analyses were used to extract features. They selected four predictors including grey matter (GM), white matter (WM), fractional anisotropy (FA) and mean diffusivity (MD). The SVM algorithm showed a relatively high accuracy for distinguishing rET from tPD when GM or WM was used as a single predictor, compared with FA and MD. Results also showed that algorithm using all the four predictors achieved 100% accuracy for distinguishing rET from tPD. In addition to studying PD patients with tremor, cognitive decline in PD patients was focused. As is known to all, different levels of cortical thinning in PD patients when mild cognitive impairment or dementia was manifest can be measured by sMRI. Morales et al. studied four classification models (naïve Bayes, multivariate filter-based naïve Bayes, filter selective naïve Bayes and SVM) to evaluate their capacity to discriminate between the three different phases of PD: cognitively intact (PDCI), with mild cognitive impairment (PDMCI), and with dementia (PDD) [70]. This analysis demonstrated that the multivariate filter-based naïve Bayes model was the best classifier, having the highest crossvalidated sensitivity, specificity and accuracy. Meanwhile, this study showed that volumes of cerebral WM and the lateral ventricles and hippocampi closely tracked cognition status and was a best predictor of dementia in PD. Feis et al. considered the asymmetry of motor symptoms of PD patients and used a multimodal method to model symptom side when PD was onset. They analyzed morphological features based on diffusion MR parameters and a multi-kernel SVM classification [71]. Remarkably, the framework yielded an accuracy of 96% in 24 PD patients. These differential diagnosis studies concerning different aspects of PD patients were based on clinical requirements and somewhat helpful in clinical decision. The differential diagnosis studies applied mature algorithms and achieved relatively high accuracies in small samples. Researchers were working hard on these themes which are fairly practical and significant in clinical practice of differential diagnosis of PD. A brief summary can be found in Table 2.

ACS Paragon Plus Environment

Page 8 of 23

Page 9 of 23 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Chemical Neuroscience

Table 2 Summery of studies using magnetic resonance imaging and artificial intelligence in studies of differential diagnosis of Parkinson's disease Papers

MRI Modalitie s

Data source

Subjects

Feature extracti on method

Validati on method

Class ifier

Acc.

Affected regions

Salvatore et al. 2014 [63]

sMRI

recruited

28 PD/28 PSP

PCA

LOOCV

SVM

88.9%

midbrain, pons, corpus callosum and thalamus

Focke et al. 2011 [64]

sMRI

recruited

21 IPS/11 MSA/10 PSP/22NC

group compari son analysis

LOOCV

SVM

IPS vs. PSP 96.8% IPS vs. MSA 71.9%

putamen, superior parietal lobe, precuneus, external capsule, corticospinal tract, precentral gyrus, occipital pole, pons, the mesencephalon, dorsal basal ganglia, cerebellar peduncles

Huppertz et al. 2016 [65]

sMRI

recruited

204PD/106 PSP/21MS AC/60MSAP/73NC

group compari son analysis

LOOCV

SVM

>80%

midbrain, basal ganglia, and cerebellar peduncles

Haller et al. 2013 [66]

SWI

recruited

16 PD /20 other Parkinsoni sm

a Relieff feature selectio n

10-fold cross validati on

SVM

86 %

thalamus and substantia nigra

Haller et al. 2012 [67]

DTI

recruited

17 PD /23 other Parkinsoni sm

a Relieff feature selectio n

10-fold cross validati on

SVM

97%

a bilateral network, predominantly in the right frontal white matter

Cherubin i et al. 2014 [68]

Multimodality: DTI, sMRI

recruited

21 PSP/ 57 PD

group compari son analysis

LOOCV

SVM

100%

basal ganglia, midbrain, cerebellum, corpus callosum

Cherubin

Multi-

recruited

15 rET/15

group

LOOCV

SVM

100%

caudate nucleus,

ACS Paragon Plus Environment

brain

ACS Chemical Neuroscience 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

i et al. 2014 [69]

modality: DTI, sMRI

Morales et al. 2013 [70]

sMRI

Feis et al. 2015 [71]

Multimodality: DTI, sMRI

Page 10 of 23

tPD

compari son analysis

globus pallidus, midbrain, internal capsule, body of the corpus callosum, and cerebellum

recruited

14 PDD/15 PDMCI/16 PDCI

FSS

5-fold cross validati on

naïve Baye s

PDD vs. PDCI 93% PDD vs. PDMC 96% PDMCI vs. PDCI 86% PDD vs. PDMCI vs. PDCI 64%

caudate, entorhinal cortical, hippocampus, brain stem, cerebellum, lateral ventricle

recruited

12 left sided and 12 rightsidedsympt om onset PD

FSS

LOOCV

SVM

96%

The right hippocampus

Acc.: Accuracy, SVM: support vector machine, PD: Parkinson's disease, PSP: progressive supranuclear palsy, MSA: multiple system atrophy, IPS: idiopathic Parkinson syndrome, NC: normal controls, LOOCV: leave-one-out cross-validation method, rsfMRI: resting state functional MRI, sMRI: structural MRI, SWI: susceptibility-weighted imaging, DTI: diffusion tensor imaging, PSA: principle component analysis-based feature selection, rET: tremor with rest tremor, tPD: tremor-dominant Parkinson's disease, PDD: Parkinson’s disease with dementia, PDMCI: Parkinson’s disease with mild cognitive impairment, PDCI: Parkinson’s disease cognitively intact, FSS: Feature subset selection.

Use of multimodal MRI in studying PD subtyping Because of the diversity of clinical manifestations of PD patients and the interaction between different clinical manifestations, some scholars believe that PD may not be a homogeneous disease, but can be divided into different subtypes. In the early 1990s, Jankovic et al. analyzed the data of a large cohort of patients and systematically proposed the classification of clinical subtypes of PD [8]. This classification was continuously improved: according to clinical symptoms, PD can be divided into TD, AR and PIGD [9]. Functional and structural differences among different subtypes should be taken into account in ML brain studies of PD using MRI. It is helpful to better treat the disease,

ACS Paragon Plus Environment

Page 11 of 23 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Chemical Neuroscience

predict prognosis in individual level, and understand the neurobiological mechanism. In subtyping of PD, the only research using MRI and machine learning we found was did by our team in 2016. We applied a SVM classifier with the recursive feature elimination method to multimodal MRI data for selecting features, and evaluated the proposed classifier with the leave-one-out cross-validation [72]. The accuracy for different subtypes of PD (19 PIGD patients and 33 nonPIGD patients) was up to 92.31%. Details were showed in Table 3. At present, our understanding of brain differences among PD subtypes is still far from comprehensive and sufficient. We believed the functional and structural MR (including T1-weighted and DTI data) differences between the subtypes must be integrated for further exploring definitive brain biomarkers specific for different subtypes. Analyzing and searching for the possible structural functional and metabolic basis of different subtypes with advanced MRI modalities and AI algorithm may be the key scientific problems that need to be solved urgently in this field. Figure 1 summarizes the abnormal changes in different brain regions in PD patients discovered by recent MRI studies. Table 3 Details of our team’s study of subtyping of Parkinson's disease Paper s

MRI Modaliti es

Data source

Subjects

Feature extractio n method

Validati on method

Classifi er

Acc.

Affected brain regions

Gu et al. 2016 [72]

Multimodality : sMRI, rs-fMRI, DTI

recruite d

52 PD (19PIG D) /45 NC

recursive feature eliminati on

LOOCV

SVM

92.31 %

frontal, parietal, occipital, temporal cortices, cerebellu m

Acc.: Accuracy, SVM: support vector machine, PD: Parkinson's disease, NC: normal controls, LOOCV: leave-one-out cross-validation method, rsfMRI: resting state functional MRI, sMRI: structural MRI, DTI: diffusion tensor imaging, PIGD: postural instability and gait difficulty.

ACS Paragon Plus Environment

ACS Chemical Neuroscience 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Figure 1 MRI can detect various abnormalities of brain regions in patients with PD and reflect the difference of disease subtypes. PD: Parkinson's disease, PSP: progressive supranuclear palsy, MSA: multiple system atrophy, fMRI: functional MRI, sMRI: structural MRI, DWI: diffusion-weighted imaging, AR: Akineto-rigid, PIGD: postural instability and gait difficulty, TD: tremor dominant.

The overall workflow of AI in recent studies Before machine learning methods based classification algorithms to build a diagnostic model with imaging features available, researchers need to process the images and compute features. So the major steps include image processing, feature extraction, feature evaluation, feature labeling and selection, training and finally classification. In recent years, some new methods, such as multi-atlas segmentation [73], texture feature application [74], multivariate fusion [75], have been applied to improve the process. But in PD workflow, most studies still followed the traditional workflow. Deep learning is currently developing at an unprecedented speed. There are two major methods of deep learning for neuroimaging studies. One has only two steps: image processing (registration and segmentation) and deep learning-based classification [76-78]. The other extracts features from the whole image information (patch) of the related region. After acquiring better advanced features, it is applied to classifier construction. So it has three steps: image processing, deep feature extraction, and classification [79]. These models are machine learning with image input or image-based machine learning methods. They skip traditional machine learning steps for feature selection but enable the entire process to map from raw input images to the final classification. We purposed that these deep learning based methods might be used in PD studies soon. Figure 2 shows a sketch map of the overall workflow in recent studies.

ACS Paragon Plus Environment

Page 12 of 23

Page 13 of 23 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Chemical Neuroscience

Figure 2 the overall workflow in building a classifier. (A) The traditional workflow; (B, C) Deep learning based workflows. ROI: region of interest, PCA: principal component analysis, SVM: support vector machine, RF: random forest.

Conclusion In this paper, we reviewed recent researches focusing on using MRI data for diagnosis PD. Based on the review; we find that there are several problems in the current time: First, most studies were based on small sizes of samples, while the accuracies of large-sample researches were generally below or barely above 80%, which were far from the standard that can be applied to clinical practice. Also, the large-scale studies were all based on PPMI dataset. Since data from one research center may be limited and over fitting, multicenter studies are preferred in the future. Second, due to the lack of multimodal MRI data, most studies were limited to single modality. The incomplete utilization of multimodal information may be an important reason for the low accuracy of the current research. Third, most studies failed to analyze combining with PD's pathological mechanism. The subjects were seldom subtyped according to the clinical symptoms or disease courses. Researchers have ignored the clinical complexity and diversity of PD patients. Fourth, classic machine learning algorithms were still the mainstream methodologies used in the existing studies. Most studies used leave-one-out or 10 fold cross-validation, which can lead to overfitting and poor generalization. The previous researches used traditional preprocess and classification techniques. In methods, there is still huge room for improvement of ROI acquirement, feature extraction, feature fusion, etc. Based on these findings, we make some recommendations on future research: First, future research should pay more attention to using multimodal MRI for diagnosis PD at the

ACS Paragon Plus Environment

ACS Chemical Neuroscience 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

prodromal stage. A growing number of evidence suggests that effective treatment slowing down or halting the progression of PD should focus on its prodromal phase [6]. Optimally, effective MRI markers and treatment should target the preclinical stage, possibly, before irreversible damages triggered by the death of dopaminergic neurons. Second, though MRI gives substantial neuronal injury biomarkers for constructing the classifier, pathologic biomarkers and quantified non-imaging parameters including genetics, serum/CSF tests, etc. of PD should be integrated along with imaging parameters in order to achieve a better diagnosis. Third, a primary challenge underlying the clinical use of PD classification models is the ability that allows good generalization to new patient data [80]. Future studies may focus on comparison of different studies using the same subjects from one dataset. Efforts should be made to standardize such comparison. Last but not least, along with the rapid development of deep learning in medicine, PD early diagnosis, differential diagnosis and subtyping using multimodal MRI can benefit from deep learning application. More research is expected in deep learning-based approaches for multimodal brain MRI image analysis, for feature selection and classification result production [79, 81-82].

Graphical Table of Contents

Author Information Department of Radiology, the Second Affiliated Hospital of Zhejiang University, School of Medicine. No.88 Jiefang Road, Shangcheng District, Hangzhou, China, 31000. Author Contributions: Jingjing Xu and Minming Zhang devised the study, the main conceptual ideas and proof outline. Jingjing Xu wrote the manuscript with support from Minming Zhang. Jingjing Xu and Minming Zhang discussed and commented on the manuscript. Funding Sources: This work was supported by the 13th Five-year Plan for National Key Research and Development Program of China (Grant No. 2016YFC1306600), the National Natural Science Foundation of China (Grant Nos. 81571654, 81371519, 81701647 and 81771820), the 12th Five-year Plan for National Science and Technology Supporting Program of China (Grant No. 2012BAI10B04), the

ACS Paragon Plus Environment

Page 14 of 23

Page 15 of 23 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Chemical Neuroscience

Fundamental Research Funds for the Central Universities of China (Grant No. 2017XZZX001-01), the Natural Science Foundation of Zhejiang Province (Grant No. Q19H180029) and the Projects of Medical and Health Technology Development Program in Zhejiang Province (Grant No. 2015KYB174). Conflict of Interest: None.

References 1. de Lau, L.-M., and Breteler, M.M. (2006) Epidemiology of Parkinson's disease. Lancet Neurol. 5(6), 525-535. 2. Zhang, Z.-X., Roman, G.-C., Hong, Z., Wu, C.-B., Qu, Q.-M., Huang, J.-B., Zhou, B., Geng, Z.-P., Wu, J.-X., Wen, H.-B., Zhao H., and Zahner, G.E. (2005) Parkinson's disease in China: prevalence in Beijing, Xian, and Shanghai. Lancet. 365(9459), 595-597. 3. de Lau, L.-M., and Breteler, M.-M. (2006) Epidemiology of Parkinson's disease. Lancet Neurol. 5(6), 525-535. 4. Postuma, R.-B., Berg, D., Stern, M., Poewe, W., Olanow, C.-W., Oertel, W., Obeso, J., Marek, K., Litvan, I., Lang, A.-E., Halliday, G., Goetz, C.-G., Gasser, T., Dubois, B., Chan, P., Bloem, B.-R., Adler C.-H., and Deuschl, G. (2015) MDS clinical diagnostic criteria for Parkinson's disease. Mov Disord. 30(12), 1591-1601. 4. Hughes, A.-J., Daniel, S.-E., Kilford L., and Lees, A.-J. (1992) Accuracy of clinical diagnosis of idiopathic Parkinson's disease: a clinico-pathological study of 100 cases. J Neurol Neurosurg Psychiatry. 55(3), 181-1844. 5. Hawkes, C.-H., Del Tredici, K., and Braak, H. (2010) A timeline for Parkinson's disease. Parkinsonism Relat Disord. 16(2), 79-84. 6. Kang, G.-A., Bronstein, J.-M., Masterman, D.-L., Redelings, M., Crum J.-A., and Ritz, B. (2005) Clinical characteristics in early Parkinson's disease in a central California population-based study. Mov Disord. 20(9), 1133-1142. 7. Zaidel, A., Arkadir, D., Israel Z., and Bergman, H. (2009) Akineto-rigid vs. tremor syndromes in Parkinsonism. Curr Opin Neurol. 22(4), 387-393. 8. Jankovic, J., McDermott, M., Carter, J., Gauthier, S., Goetz, C., Golbe, L., Huber, S., Koller, W., Olanow, C., Shoulson I., and et al. (1990) Variable expression of Parkinson's disease: a base-line analysis of the DATATOP cohort. The Parkinson Study Group. Neurology. 40(10), 1529-1534. 9. Post, B., Merkus, M.-P., de Haan, R.-J., Speelman J.-D., and Group, C.-S. (2007) Prognostic factors for the progression of Parkinson's disease: a systematic review. Mov Disord. 22(13), 18391851; quiz 1988. 10. Lewis, S.-J., Foltynie, T., Blackwell, A.-D., Robbins, T.-W., Owen A.-M., and Barker, R.-A. (2005) Heterogeneity of Parkinson's disease in the early clinical stages using a data driven approach. J Neurol Neurosurg Psychiatry. 76(3), 343-348. 11. Stefanova, N., Bucke, P., Duerr S., and Wenning, G.-K. (2009) Multiple system atrophy: an update. Lancet Neurol. 8(12), 1172-1178. 12. Williams, D.-R. and Lees, A.-J. (2009) Progressive supranuclear palsy: clinicopathological concepts and diagnostic challenges. Lancet Neurol. 8(3), 270-279.

ACS Paragon Plus Environment

ACS Chemical Neuroscience 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

13. Rizzo, G., Copetti, M., Arcuti, S., Martino, D., Fontana A., and Logroscino, G. (2016) Accuracy of clinical diagnosis of Parkinson disease: A systematic review and meta-analysis. Neurology. 86(6), 566-576. 14. DeKosky, S.-T. and Marek, K. (2003) Looking backward to move forward: early detection of neurodegenerative disorders. Science. 302(5646), 830-834. 15. Damier, P., Hirsch, E.-C., Agid Y., and Graybiel, A.-M. (1999) The substantia nigra of the human brain. II. Patterns of loss of dopamine-containing neurons in Parkinson's disease. Brain. 122 (Pt 8), 1437-1448. 16. Obeso, J.-A., Rodriguez-Oroz, M.-C., Rodriguez, M., Lanciego, J.-L., Artieda, J., Gonzalo N., and Olanow, C.-W. (2000) Pathophysiology of the basal ganglia in Parkinson's disease. Trends Neurosci. 23(10 Suppl), S8-19. 17. Seppi, K. and Poewe, W. (2010) Brain magnetic resonance imaging techniques in the diagnosis of parkinsonian syndromes. Neuroimaging Clin N Am. 20(1), 29-55. 18. Mahlknecht, P., Hotter, A., Hussl, A., Esterhammer, R., Schocke, M., and Seppi, K. (2010) Significance of MRI in diagnosis and differential diagnosis of Parkinson's disease. Neurodegener Dis. 7(5), 300-318. 19. Beyer, M.-K., Janvin, C.-C., Larsen, J.-P., and Aarsland, D. (2007) A magnetic resonance imaging study of patients with Parkinson’s disease with mild cognitive impairment and dementia using voxel-based morphometry. J Neurol Neurosurg Psychiatry.78 (3), 254-259. 20. Tessa, C., Lucetti, C., Giannelli, M., Diciotti, S., Poletti, M., Danti, S., Baldacci, F., Vignali, C., Bonuccelli, U., Mascalchi, M., and Toschi, N. (2014) Progression of brain atrophy in the early stages of Parkinson’s disease: A longitudinal tensor-based morphometry study in de novo patients without cognitive impairment. Hum Brain Mapp. 35 (8), 3932-3944. 21. Guan, X., Zeng, Q., Guo, T., Wang, J., Xuan, M., Gu, Q., Wang, T., Huang, P., Xu, X., and Zhang, M. (2017) Disrupted Functional Connectivity of Basal Ganglia across Tremor-Dominant and Akinetic/Rigid-Dominant Parkinson's Disease. Front Aging Neurosci. 9, 360. 22. Nagae, L.-M., Honce, J.-M., Tanabe, J., Shelton, E., Sillau, S.-H.,and Berman, B.-D. (2016) Microstructural Changes within the Basal Ganglia Differ between Parkinson Disease Subtypes. Front Neuroanat. 10, 17. 23. Fearnley, J.-M. and Lees, A.-J. (1991) Ageing and Parkinson's disease: substantia nigra regional selectivity. Brain. 114 ( Pt 5), 2283-2301. 24. Damier, P., Hirsch, E.C., Agid Y., and Graybiel, A.-M. (1999) The substantia nigra of the human brain. II. Patterns of loss of dopamine-containing neurons in Parkinson's disease. Brain. 122 ( Pt 8), 1437-1448. 25. Guan, X., Xuan, M., Gu, Q., Huang, P., Liu, C., Wang, N., Xu, X., Luo, W., and Zhang, M. (2017) Regionally progressive accumulation of iron in Parkinson's disease as measured by quantitative susceptibility mapping. NMR Biomed. 30(4). 26. Guan, X., Xuan, M., Gu, Q., Xu, X., Huang, P., Wang, N., Shen, Z., Xu, J., Luo, W., and Zhang, M. (2017) Influence of regional iron on the motor impairments of Parkinson's disease: A quantitative susceptibility mapping study. J Magn Reson Imaging. 45(5),1335-1342. 27. Guan, X., Xu, X., and Zhang, M. (2017) Region-Specific Iron Measured by MRI as a Biomarker for Parkinson's Disease. Neurosci Bull. 33(5), 561-567. 28. Xuan, M., Guan, X., Gu, Q., Shen, Z., Yu, X., Qiu, T., Luo, X., Song, R., Jiaerken, Y., Xu, X., Huang, P., Luo, W., and Zhang, M. (2017) Different iron deposition patterns in early- and middle-

ACS Paragon Plus Environment

Page 16 of 23

Page 17 of 23 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Chemical Neuroscience

late-onset Parkinson's disease. Parkinsonism Relat Disord. 44, 23-27. 29. Zeng, Q., Guan, X., Law Yan Lun, JCF., Shen, Z., Guo, T., Xuan, M., Gu, Q., Xu, X., Chen, M., and Zhang, M. (2017) Longitudinal Alterations of Local Spontaneous Brain Activity in Parkinson's Disease. Neurosci Bull. 33(5),501-509. 30. Gu, Q., Cao, H., Xuan, M., Luo, W., Guan, X., Xu, J., Huang, P., Zhang, M., and Xu, X. (2016) Increased thalamic centrality and putamen-thalamic connectivity in patients with parkinsonian resting tremor. Brain Behav. 7(1),e00601. 31. Xu, X., Guan, X., Guo, T., Zeng, Q., Ye, R., Wang, J., Zhong, J., Xuan, M., Gu, Q., Huang, P., Pu, J., Zhang, B., and Zhang, M. (2018) Brain Atrophy and Reorganization of Structural Network in Parkinson's Disease With Hemiparkinsonism. Front Hum Neurosci. 12,117. 32. Xuan, M., Guan, X., Huang, P., Shen, Z., Gu, Q., Yu, X., Xu, X., Luo, W., and Zhang, M. (2017) Different patterns of gray matter density in early- and middle-late-onset Parkinson's disease: a voxel-based morphometry study. Brain Imaging Behav. 30. 33. Guan, X., Huang, P., Zeng, Q., Liu, C., Wei, H., Xuan, M., Gu, Q., Xu, X., Wang, N., Yu, X., Luo, X., and Zhang, M. (2018) Quantitative susceptibility mapping as a biomarker for evaluating white matter alterations in Parkinson's disease. Brain Imaging Behav. published online Feb 7, 2018. DOI: 10.1007/s11682-018-9842-z. 34. Tang, Y., Meng, L., Wan, C.-M., Liu, Z.-H., Liao, W.-H., Yan, X.-X., Wang, X.-Y., Tang, B.S., and Guo, J.-F. (2017) Identifying the presence of Parkinson's disease using low-frequency fluctuations in BOLD signals. Neurosci Lett. 645,1-6. 35. Blazejewska, A.-I., Schwarz, S.-T., Pitiot, A., Stephenson, M.-C., Lowe, J., Bajaj, N., Bowtell, R.-W., Auer, D.-P., and Gowland, P.-A. (2013) Visualization of nigrosome 1 and its loss in PD: pathoanatomical correlation and in vivo 7 T MRI. Neurology. 81(6), 534-540. 36. Reiter, E., Mueller, C., Pinter, B., Krismer, F., Scherfler, C., Esterhammer, R., Kremser, C., Schocke, M., Wenning, G.-K., Poewe, W., and Seppi, K. (2015) Dorsolateral nigral hyperintensity on 3.0T susceptibility-weighted imaging in neurodegenerative Parkinsonism. Mov Disord. 30(8), 1068-1076. 37. Schwarz, S.-T., Afzal, M., Morgan, P.-S., Bajaj, N., Gowland, P.-A., and Auer, D.P. (2014) The 'swallow tail' appearance of the healthy nigrosome-a new accurate test of Parkinson's disease: a case-control and retrospective cross-sectional MRI study at 3T. PLoS One. 9(4), e93814. 38. Sasaki, M., Shibata, E., Tohyama, K., Takahashi, J., Otsuka, K., Tsuchiya, K., Takahashi, S., Ehara, S., Terayama, Y., and Sakai, A. (2006) Neuromelanin magnetic resonance imaging of locus ceruleus and substantia nigra in Parkinson's disease. Neuroreport. 17(11), 1215-1218. 39. Tang, Y., Meng, L., Wan, C.-M., Liu, Z.-H., Liao, W.-H., Yan, X.-X., Wang, X.-Y., Tang, B.S., and Guo, J.-F. (2017) Identifying the presence of Parkinson's disease using low-frequency fluctuations in BOLD signals. Neurosci Lett. 645,1-6. 40. Long, D., Wang, J., Xuan, M., Gu, Q., Xu, X., Kong, D., and Zhang, M. (2012) Automatic classification of early Parkinson's disease with multi-modal MR imaging. PLoS One. 7(11), e47714. 41. Chen, Y., Storrs, J., Tan, L., Mazlack, L.-J., Lee, J.-H., and Lu, L.-J. (2014) Detecting brain structural changes as biomarker from magnetic resonance images using a local feature based SVM approach. J Neurosci Methods. 221,22-31. 42. Chen, Y., Yang, W., Long, J., Zhang, Y., Feng, J., Li, Y., and Huang, B. (2015) Discriminative analysis of Parkinson's disease based on whole-brain functional connectivity. PLoS One. 10(4), e0124153.

ACS Paragon Plus Environment

ACS Chemical Neuroscience 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

43. Adeli, E., Shi, F., An, L., Wee, C.-Y., Wu, G., Wang, T., and Shen, D. (2016) Joint featuresample selection and robust diagnosis of Parkinson's disease from MRI data. Neuroimage. 141,206219. 44. Adeli, E., Wu, G., Saghafi, B., An, L., Shi, F., and Shen, D. (2017) Kernel-based Joint Feature Selection and Max-Margin Classification for Early Diagnosis of Parkinson's Disease. Sci Rep. 7, 41069. 45. Pläschke, R.-N., Cieslik, E.-C., Müller, V.-I., Hoffstaedter, F., Plachti, A., Varikuti, D.-P., Goosses, M., Latz, A., Caspers, S., Jockwitz, C., Moebus, S., Gruber, O., Eickhoff, C.-R., Reetz, K., Heller, J., Südmeyer, M., Mathys, C., Caspers, J., Grefkes, C., Kalenscher, T., Langner, R., and Eickhoff, S.-B. (2017) On the integrity of functional brain networks in schizophrenia, Parkinson's disease, and advanced age: Evidence from connectivity-based single-subject classification. Hum Brain Mapp. 38(12), 5845-5858. 46. Peng, B., Wang, S., Zhou, Z., Liu, Y., Tong, B., Zhang, T., and Dai, Y. (2017) A multilevelROI-features-based machine learning method for detection of morphometric biomarkers in Parkinson's disease. Neurosci Lett. 651, 88-94. 47. Dinov, I.-D., Heavner, B., Tang, M., Glusman, G., Chard, K., Darcy, M., Madduri, R., Pa, J., Spino, C., Kesselman, C., Foster, I., Deutsch, E.-W., Price, N.-D., Van Horn, J.-D., Ames, J., Clark, K., Hood, L., Hampstead, B.-M., Dauer, W., and Toga, A.-W. (2016) Predictive Big Data Analytics: A Study of Parkinson's Disease Using Large, Complex, Heterogeneous, Incongruent, Multi-Source and Incomplete Observations. PLoS One. 11(8), e0157077. 48. Amoroso, N., La Rocca, M., Monaco, A., Bellotti, R., and Tangaro, S. (2018) Complex networks reveal early MRI markers of Parkinson's disease. Med Image Anal. 48,12-24. 49. Liu, H., Du, G., Zhang, L., Lewis, M.-M., Wang, X., Yao, T., Li, R., and Huang, X. (2016) Folded concave penalized learning in identifying multimodal MRI marker for Parkinson's disease. J Neurosci Methods. 268, 1-6. 50. Zeng, L.-L., Xie, L., Shen, H., Luo, Z., Fang, P., Hou, Y., Tang, B., Wu, T., and Hu, D. (2017) Differentiating Patients with Parkinson's Disease from Normal Controls Using Gray Matter in the Cerebellum. Cerebellum. 16(1), 151-157. 51. Ramli, N., Nair, S.-R., Ramli, N.-M., and Lim, S.Y. (2015) Differentiating multiple-system atrophy from Parkinson's disease. Clin Radiol. 70(5), 555-564. 52. Nair, S.-R., Tan, L.-K., Mohd Ramli, N., Lim, S.-Y., Rahmat, K., and Mohd Nor, H. (2013) A decision tree for differentiating multiple system atrophy from Parkinson's disease using 3-T MR imaging. Eur Radiol. 23(6), 1459-1466. 53. Messina, D., Cerasa, A., Condino, F., Arabia, G., Novellino, F., Nicoletti, G., Salsone, M., Morelli, M., Lanza, P.-L., andQuattrone, A. (2011) Patterns of brain atrophy in Parkinson's disease, progressive supranuclear palsy and multiple system atrophy. Parkinsonism Relat Disord. 17(3), 172-176. 54. Ito, M., Watanabe, H., Kawai, Y., Atsuta, N., Tanaka, F., Naganawa, S., Fukatsu, H., and Sobue, G. (2007) Usefulness of combined fractional anisotropy and apparent diffusion coefficient values for detection of involvement in multiple system atrophy. J Neurol Neurosurg Psychiatry. 78(7), 722-728. 55. Pellecchia, M.-T., Barone, P., Mollica, C., Salvatore, E., Ianniciello, M., Longo, K., Varrone, A., Vicidomini, C., Picillo, M., De Michele, G., Filla, A., Salvatore, M., and Pappata, S. (2009) Diffusion-weighted imaging in multiple system atrophy: a comparison between clinical subtypes.

ACS Paragon Plus Environment

Page 18 of 23

Page 19 of 23 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Chemical Neuroscience

Mov Disord. 24(5), 689-696. 56. Yao, Q., Zhu, D., Li, F., Xiao, C., Lin, X., Huang, Q., and Shi, J. (2017) Altered Functional and Causal Connectivity of Cerebello-Cortical Circuits between Multiple System Atrophy (Parkinsonian Type) and Parkinson's Disease. Front Aging Neurosci. 9, 266. 57. Massey, L.-A., Micallef, C., Paviour, D.-C., O'Sullivan, S.-S., Ling, H., Williams, D.-R., Kallis, C., Holton, J.-L., Revesz, T., Burn, D.-J., Yousry, T., Lees, A.-J., Fox, N.-C., and Jager, H.R. (2012) Conventional magnetic resonance imaging in confirmed progressive supranuclear palsy and multiple system atrophy. Mov Disord. 27(14), 1754-1762. 58. Nicoletti, G., Tonon, C., Lodi, R., Condino, F., Manners, D., Malucelli, E., Morelli, M., Novellino, F., Paglionico, S., Lanza, P., Messina, D., Barone, P., Morgante, L., Zappia, M., Barbiroli, B., and Quattrone, A. (2008) Apparent diffusion coefficient of the superior cerebellar peduncle differentiates progressive supranuclear palsy from Parkinson's disease. Mov Disord. 23(16), 23702376. 59. Paviour, D.-C., Thornton, J.-S., Lees, A.-J., and Jager, H.R. (2007) Diffusion-weighted magnetic resonance imaging differentiates Parkinsonian variant of multiple-system atrophy from progressive supranuclear palsy. Mov Disord. 22(1), 68-74. 60. Hingwala, D.-R., Kesavadas, C., Thomas, B., and Kapilamoorthy, T.-R. (2013) Susceptibility weighted imaging in the evaluation of movement disorders. Clin Radiol. 68(6), e338-48. 61. Vymazal, J., Righini, A., Brooks, R.-A., Canesi, M., Mariani, C., Leonardi, M., and Pezzoli, G. (1999) T1 and T2 in the brain of healthy subjects, patients with Parkinson disease, and patients with multiple system atrophy: relation to iron content. Radiology. 211(2), 489-495. 62. Wang, Y., Butros, S.-R., Shuai, X., Dai, Y., Chen, C., Liu, M., Haacke, E.-M., Hu, J., and Xu, H. (2012) Different iron-deposition patterns of multiple system atrophy with predominant parkinsonism and idiopathetic Parkinsondiseases demonstrated by phase-corrected susceptibilityweighted imaging. AJNR Am J Neuroradiol. 33(2), 266-273. 63. Salvatore, C., Cerasa, A., Castiglioni, I., Gallivanone, F., Augimeri, A., Lopez, M., Arabia, G., Morelli, M., Gilardi, M.-C., and Quattrone, A. (2014) Machine learning on brain MRI data for differential diagnosis of Parkinson's disease and Progressive Supranuclear Palsy. J Neurosci Methods. 222, 230-237. 64. Focke, N.-K., Helms, G., Scheewe, S., Pantel, P.-M., Bachmann, C.-G., Dechent, P., Ebentheuer, J., Mohr, A., Paulus, W., and Trenkwalder, C. (2011) Individual voxel-based subtype prediction can differentiate progressive supranuclear palsy from idiopathic Parkinson syndrome and healthy controls. Hum Brain Mapp. 32(11),1905-1915. 65. Huppertz, H.-J., Möller, L., Südmeyer, M., Hilker, R., Hattingen, E., Egger, K., Amtage, F., Respondek, G., Stamelou, M., Schnitzler, A., Pinkhardt, E.-H., Oertel, W.-H., Knake, S., Kassubek, J., and Höglinger, G.U. (2016) Differentiation of neurodegenerative parkinsonian syndromes by volumetric magnetic resonance imaging analysis and support vector machine classification. Mov Disord. 31(10),1506-1517. 66. Haller, S., Badoud, S., Nguyen, D., Barnaure, I., Montandon, M.-L., Lovblad, K.-O., and Burkhard, P.-R. (2013) Differentiation between Parkinson disease and other forms of Parkinsonism using support vector machine analysis of susceptibility-weighted imaging (SWI): initial results. Eur Radiol. 23(1), 12-9. 67. Haller, S., Badoud, S., Nguyen, D., Garibotto, V., Lovblad, K.-O., and Burkhard, P.-R. (2012) Individual detection of patients with Parkinson disease using support vector machine analysis of

ACS Paragon Plus Environment

ACS Chemical Neuroscience 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

diffusion tensor imaging data: initial results. AJNR Am J Neuroradiol. 33(11), 2123-2128. 68. Cherubini, A., Morelli, M., Nisticó, R., Salsone, M., Arabia, G., Vasta, R., Augimeri, A., Caligiuri, M.-E., and Quattrone, A. (2014) Magnetic resonance support vector machine discriminates between Parkinson disease and progressive supranuclear palsy. Mov Disord. 29(2), 266-269. 69. Cherubini, A., Nisticó, R., Novellino, F., Salsone, M., Nigro, S., Donzuso, G., and Quattrone, A. (2014) Magnetic resonance support vector machine discriminates essential tremor with rest tremor from tremor-dominant Parkinson disease. Mov Disord. 29(9), 1216-1219. 70. Morales, D.-A., Vives-Gilabert, Y., Gómez-Ansón, B., Bengoetxea, E., Larrañaga, P., Bielza, C., Pagonabarraga, J., Kulisevsky, J., Corcuera-Solano, I., and Delfino, M. (2013) Predicting dementia development in Parkinson's disease using Bayesian network classifiers. Psychiatry Res. 213(2), 92-98. 71. Feis, D.-L., Pelzer, E.-A., Timmermann, L., and Tittgemeyer, M. (2015) Classification of symptom-side predominance in idiopathic Parkinson's disease. NPJ Parkinsons Dis. 1, 15018. 72. Gu, Q., Zhang, H., Xuan, M., Luo, W., Huang, P., Xia, S., and Zhang, M. (2016) Automatic Classification on Multi-Modal MRI Data for Diagnosis of the Postural Instability and Gait Difficulty Subtype of Parkinson's Disease. J Parkinsons Dis. 6(3), 545-556. 73. Wang, H., Suh, J.-W., Das, S.-R., Pluta, J.-B., Craige, C., and Yushkevich, P.A. (2013) Multiatlas segmentation with joint label fusion. IEEE Trans Pattern Anal Mach Intell. 35(3), 611-623. 74. Zacharaki, E.-I., Wang, S., Chawla, S., Soo Yoo, D., Wolf, R., Melhem, E.-R., and Davatzikos, C. (2009) Classification of brain tumor type and grade using MRI texture and shape in a machine learning scheme. Magn Reson Med. 62(6), 1609-1618. 75. Ding, X., Yang, Y., Stein, E.-A., and Ross, T.J. (2015) Multivariate classification of smokers and nonsmokers using SVM-RFE on structural MRI images. Hum Brain Mapp. 36(12), 4869-4879. 76. Sarraf, S., and Tofighi, G. (2016) DeepAD: Alzheimer’s Disease Classification via Deep Convolutional Neural Networks using MRI and fMRI. bioRxiv. 070441. 77. Liu, S., Liu, S., Cai, W., Che, H., Pujol, S., Kikinis, R., Feng, D., and Fulham, M.-J. (2015) Multimodal neuroimaging feature learning for multiclass diagnosis of Alzheimer's disease. IEEE Trans Biomed Eng. 62(4), 1132-1140. 78. Shi, J., Zheng, X., Li, Y., Zhang, Q., and Ying, S. (2018) Multimodal neuroimaging feature learning with multimodal stacked deep polynomial networks for diagnosis of Alzheimer's disease. IEEE J Biomed Health Inform. 22(1), 173-183. 79. Suk, H.-I., Lee, S.-W., Shen, D., and A.s.D.N. (2014) Initiative, Hierarchical feature representation and multimodal fusion with deep learning for AD/MCI diagnosis. NeuroImage. 101, 569-582. 80. Akkus, Z., Galimzianova, A., Hoogi, A., Rubin, D.-L., and Erickson, B.-J. (2017) Deep learning for brain MRI segmentation: state of the art and future directions. J Digit Imaging. 30(4), 449-459. 81. Dolz, J., Desrosiers, C., and Ayed, I.-B. (2018) 3D fully convolutional networks for subcortical segmentation in MRI: A large-scale study. NeuroImage. 170, 456-470. 82. Liu, X, Chen, K, Wu, T, Weidman, D, Lure, F, and Li, J. (2018) Use of multimodality imaging and artificial intelligence for diagnosis and prognosis of early stages of Alzheimer's disease. Transl Res. 194, 56-67.

ACS Paragon Plus Environment

Page 20 of 23

Page 21 of 23 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Chemical Neuroscience

Figure 1 MRI can detect various abnormalities of brain regions in patients with PD and reflect the difference of disease subtypes. PD: Parkinson's disease, PSP: progressive supranuclear palsy, MSA: multiple system atrophy, fMRI: functional MRI, sMRI: structural MRI, DWI: diffusion-weighted imaging, AR: Akineto-rigid, PIGD: postural instability and gait difficulty, TD: tremor dominant.

ACS Paragon Plus Environment

ACS Chemical Neuroscience 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Figure 2 the overall workflow in building a classifier. (A) The traditional workflow; (B, C) Deep learning based workflows. ROI: region of interest, PCA: principal component analysis, SVM: support vector machine, RF: random forest. 232x127mm (300 x 300 DPI)

ACS Paragon Plus Environment

Page 22 of 23

Page 23 of 23 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Chemical Neuroscience

Graphical Table of Contents

ACS Paragon Plus Environment