Determination of Imaging Biomarkers to Decipher Disease Trajectories and Differential Diagnosis of Neurodegenerative Diseases (DIsease TreND)

Journal of Neuroscience Methods xxx (xxxx) xxx–xxx

Contents lists available at ScienceDirect

Journal of Neuroscience Methods journal homepage: www.elsevier.com/locate/jneumeth

Determination of Imaging Biomarkers to Decipher Disease Trajectories and Differential Diagnosis of Neurodegenerative Diseases (DIsease TreND) ⁎

Gurpreet Singha,b,1, Lakshminarayanan Samavedhama,c, , Erle Chuen-Hian Limd, , the Alzheimer’s Disease Neuroimaging Initiative2, the Parkinson Progression Marker Initiative2 a

Department of Chemical and Biomolecular Engineering, National University of Singapore, Singapore Department of Radiology, Dalio Institute of Cardiovascular Imaging, NewYork-Presbyterian Hospital and the Weill Cornell Medicine, New York, United States c Residential College 4, 8 College Avenue West, #02-16W, Education Resource Centre, Singapore 138608, Singapore d Department of Neurology, National University Health System, National University of Singapore, Singapore b

G R A P H I C A L A B S T R A C T

A R T I C LE I N FO

A B S T R A C T

Keywords: Neurodegenerative disease diagnosis Computer-assisted diagnosis Differential diagnosis Imaging biomarkers Machine learning Unsupervised learning

Background: Understanding disease progression of neurodegenerative diseases (NDs) is important for better prognosis and decisions on the appropriate course of treatment to slow down the disease progression. New Method: We present here an innovative machine learning framework capable of (1) indicating the trajectory of disease progression by identifying relevant imaging biomarkers and (2) automated disease diagnosis. SelfOrganizing Maps (SOM) have been used for data dimensionality reduction and to reveal potentially useful disease-specific biomarkers, regions of interest (ROIs). These ROIs have been used for automated disease diagnosis using Least Square Support Vector Machines (LS-SVM) and to delineate disease progression. Results: A multi-site, multi-scanner dataset containing 1316 MRIs was obtained from ADNI3 and PPMI. Identified biomarkers have been used to decipher (1) trajectory of disease progression and (2) identify clinically relevant ROIs. Furthermore, we have obtained a classification accuracy of 94.29 ± 0.08% and 95.37 ± 0.02% for

Abbreviations: AN, anterior nucleus of thalamus; BD18, Brodmann area 18; BD19, Brodmann area 19; BD23, Brodmann area 23; BD27, Brodmann area 27; BD29, Brodmann area 29; BD30, Brodmann area 30; BD33, Brodmann area 33; CB, caudate body; CL, cerebellar lingual; CoV, culmen of vermis; DV, Declive; FT, fastigium; HT, hypothalamus; LPN, lateral posterior nucleus; MB, mammillary Body; MDN, medial dorsal nucleus; MN, midline nucleus; NoV, nodule of vermis; PV, pulvinar; RN, red nucleus; SN, substantia nigra; UoV, uvula of vermis; VAN, ventral anterior nucleus; VLN, ventral lateral nucleus; VPMN, ventral posterior medial nucleus; AD, Alzheimer disease; PD, Parkinson disease; MCI, mild cognitive impairment; SWEDD, patients with scans without evidence of dopaminergic deficit; HC, healthy controls ⁎ Corresponding author at: Department of Chemical and Biomolecular Engineering, National University of Singapore, Singapore. E-mail addresses: [email protected] (G. Singh), [email protected] (L. Samavedham), [email protected] (E.C.-H. Lim). 1 Presently working as Cognitive Software Engineer at Dalio Institute of Cardiovascular Imaging, NewYork-Presbyterian Hospital and the Weill Cornell Medicine, New York. 2 Data collection and sharing for this study was funded by the Alzheimer’s disease Neuroimaging Initiative (http://adni.loni.usc.edu/) and the Parkinson’s Progression Markers Initiative (http://www.ppmi-info.org/) primarily supported by National Institute on Aging, the National Institute of Biomedical Imaging & Bio-engineering and by the Michael J. Fox Foundation for Parkinson’s Research respectively. The investigators within the ADNI and PPMI only provided the data but they did not contribute towards analysis or writing of this report. 3 ADNI: Alzheimer’s disease neuroimaging initiative and PPMI: Parkinson’s progression markersinitiative. https://doi.org/10.1016/j.jneumeth.2018.05.009 Received 15 June 2017; Received in revised form 30 January 2018; Accepted 14 May 2018 0165-0270/ © 2018 Elsevier B.V. All rights reserved.

Please cite this article as: Singh, G., Journal of Neuroscience Methods (2018), https://doi.org/10.1016/j.jneumeth.2018.05.009

Journal of Neuroscience Methods xxx (xxxx) xxx–xxx

G. Singh et al.

distinguishing AD and PD from HC subjects respectively. Comparison with other existing methods: The goal of this study was fundamentally different from other machine learning based studies for automated disease diagnosis. We aimed to develop a method that has two-fold benefits (1) It can be used to understand pathology of neurodegenerative diseases and (2) It also achieves automated disease diagnosis. Conclusions: In the absence of established disease biomarkers, clinical diagnosis is heavily prone to misdiagnosis. Being clinically relevant and readily adaptable in the current clinical settings, the developed framework could be a stepping stone to make machine learning based Clinical Decision Support System (CDSS) for neurodegenerative disease diagnosis a reality.

1. Introduction

On MRIs, changes in the brain appear as the change in RGB color channel values. However, the location and amount of this change are localized and small as compared to the overall information present on MRIs. Thus, the need for a technique that can not only do feature extraction but also preserve the topological representation. For feature extraction, we have made use of self-organizing maps (SOM). SOM is an unsupervised learning based machine learning algorithm for non-linear dimensionality reduction well suited for large multi-dimensional datasets (Kohonen, 1998), like brain MRIs. It produces a representation of patterns in the input data by mapping it to an output space such that the nodes in the output map represent physical spatial patterns that can directly be reversibly mapped to input space unlike traditional methods such as PCA. Further, it is inherently suited for generating a low-dimensional representation of high-dimensional datasets while still preserving the inherent non-linearity and underlying distribution of data. This introduction of nonlinearity not only provides a generalized pattern of the data but also improves the topographical optimization and prevents overfitting to the data, making it less prone to noise in the data. Many brain regions affected by NDs appear as signal intensity abnormalities on MRIs. The presence of signal intensity abnormalities in a specific pattern or at a specific location is indicative of a kind of disease. However, Machine learning studies have been conducted in an isolated manner focused only on one of the two major neurodegenerative diseases at a time i.e. either AD or PD. There is an increasing evidence that NDs exhibit overlapping clinical symptoms during early stages of NDs, such as AD with subjects suffering from Mild Cognitive Impairment (MCI) (Walker et al., 2015; Xie et al., 2014). The lack of established biomarkers and similar clinical presentation makes it difficult to diagnose NDs at early stages of the disease in a clinical setting. Studying these diseases together is important to understand their epidemiology and also to assess the generalizability of a methodology. Selection of clinically relevant ROIs from these images, that can help to differentiate between two diseased subjects or a diseased subject from a healthy control, could be the key towards designing more practical CDSS. Availability of such a CDSS capable of disease prognostication by comparing the brain areas affected by disease trajectory bio-markers could greatly improve the clinical efficiency by assisting in disease prognosis. Thus, from practical perspective, it is imperative to have a CDSS capable of

The aging population is a major global demographical trend with great social, political and economic consequences. In the next four decades, the share of people aged 60 and above is expected to rise to 22% of the total population, a jump from 841 million to 2.1 billion people (United Nations, 2013). This phenomenon is closely linked to an increased prevalence of dementia. Alzheimer disease (AD) is the leading cause of dementia in the aged population, whereas Parkinson disease (PD) is the fifth most common form of dementia and one of the leading cause for movement related disorder. Both of these NDs are expected to double in numbers over the next two decades (Dorsey et al., 2007; Prince et al., 2013). Currently, there is no cure for NDs. Over the last decade, there has been a growing interest in machine learning based approaches amongst the neuroimaging community. Using machine learning, classifier models are built by extracting information from imaging modalities such as functional Magnetic Resonance Images (fMRI) (Chen et al., 2011; Kenny et al., 2012; Li et al., 2014), Positron Emission Topography (PET) (Bailly et al., 2015; Nagano-Saito et al., 2004), Diffusion Tensor Imaging (DTI) (Graña et al., 2011), structural Magnetic Resonance Imaging (MRI) (Salvatore et al., 2014; Singh and Lakshminarayanan, 2015) and others. Of these modalities, structural MRI has emerged to be a useful indicator of disease progression, is noninvasive, widely available, and have the potential to assist with clinical diagnosis and monitoring the progression of disease in most, if not all, of the NDs (Jack et al., 2010; Kassubek and Müller, 2016). Machine learning based methods are typically suited for handling high dimensional datasets, like neuroimaging modalities. Machine learning tools and statistical methods such as Principal Component Analysis (PCA) (Salvatore et al., 2014), Independent Component Analysis (ICA) (Yang et al., 2014) and Pearson's correlation (Graña et al., 2011) have been used for feature extraction. For subject classification, these features may be presented to classification algorithms such as Support Vector Machines (SVM). By combining neuroimaging modalities with machine learning algorithms, a new era of intelligent nonknowledge based clinical decision support systems (CDSS) is emerging. However, studies for automated disease diagnosis are rarely designed to provide clinically interpretable insights. Determination of reliable biomarker(s), that could be used to identify and track the progression of neurodegenerative disease(s), could be a better target for devising clinically applicable machine learning tools for neuroimaging based disease diagnosis as they could also help in understanding disease pathology. Our devised methodology has two-fold benefits wherein a unique combination of unsupervised and supervised learning algorithms has been used

1 Detecting changes in the brain tissue 2 Understanding common disease progression pathways and 3 Classifying patients in terms of predefined patient classes To achieve this, we have used machine learning based approach to identify relevant regions for differential diagnosis of the AD, PD, Mild Cognitive Impairment (MCI), Scans Without Evidence of Dopaminergic Deficit (SWEDD) and Healthy Control (HC) subjects. We have developed an innovative framework wherein unsupervised learning based Self-Organising Maps (SOM) has been used to determine unknown, but the potentially useful structure in the imaging data, Regions of Interest

1 To identify ROIs that correspond to regions known to be affected by a disease, imaging biomarker, without the availability of a priori information and 2 To obtain high classification accuracy on the largest dataset reported in the literature for machine learning based approaches for diagnosis of NDs

2

Journal of Neuroscience Methods xxx (xxxx) xxx–xxx

G. Singh et al.

2. Materials and methods

(ROIs), that could be useful to identify disease-specific biomarkers. Understanding these changes in the form of a disease trajectory is important to improve disease prognostication. Welch-Aspin Test (WAT) was used to rank selected ROIs according to their clinical relevance based on statistical inference obtained from the neuroimaging dataset. Thereafter, Least Square Support Vector Machine (LSSVM) has been used to classify subjects. Biomarkers identified in different age-group of the subjects were used to understand disease-specific trajectories. This framework has been thoroughly tested using 1316 MRIs obtained from two clinical repositories, ADNI and PPMI to determine

2.1. Data acquisition We have obtained morphological T1-weighted Magnetic Resonance Images (MRIs) for the AD, MCI, PD, SWEDD and HC subjects acquired from multiple sites and from multiple scanners at field strengths of 1.5 and 3 T from ADNI and PPMI database (Jack et al., 2008; Marek et al., 2011). Table 1 shows the demographic and clinical details of the subjects that comprise the dataset used in this work. Geriatric Depression Scale (GDS) was used to assess depressive symptoms for all the subjects. Unified Parkinson's Disease Rating Scale (MDS-UPDRS) and the Hoehn and Yahr Scale (H&Y) were used for assessing the severity of PD symptoms. The Montreal Cognitive Assessment (MoCA) and Benton Judgment of Line Orientation (Bjlo) test were used for cognitive assessment for subjects obtained from PPMI. For subjects obtained from ADNI, Mini-Mental State Exam (MMSE) and Modified Hachinski Ischemia Scale (MHIS) were used to measure cognitive impairment and

1 Population-wise indicative biomarkers for understanding the disease pathology 2 Age-related trajectory of brain areas affected in NDs to help with disease prognosis 3 It's suitability for automated differential disease diagnosis

Table 1 Demographic and clinical details of subjects. Class

AD HC MCI PD SWEDD AD HC MCI PD SWEDD AD HC MCI PD SWEDD AD HC MCI PD SWEDD AD HC MCI PD SWEDD AD HC MCI HC PD SWEDD HC PD SWEDD HC PD SWEDD HC PD SWEDD AD HC MCI

Variables

N

M/F

Age

Education

GDS

MMSE

MoCA

Bjol

MDS-UPDRS

H&Y

MHIS

AIG

ADG

51-80

51-60

61-70

71-80

128 262 447 408 71 62/66 130/132 270/177 262/146 49/22 67.30 ± 69.55 ± 69.65 ± 64.00 ± 64.59 ± 16.24 ± 16.45 ± 16.23 ± 15.16 ± 14.61 ± 05.91 ± 05.20 ± 05.59 ± 05.41 ± 05.68 ± 22.53 ± 28.89 ± 28.14 ± 28.24 ± 26.59 ± 25.94 ± 26.72 ± 24.87 ± 24.72 ± 02.90 ± 26.89 ± 21.45 ± 00.02 ± 02.02 ± 01.58 ± 00.59 ± 00.44 ± 00.50 ±

53 45 22 130 23 18/35 22/23 13-Sep 61/69 15/08 57.96 ± 55.80 ± 58.82 ± 55.47 ± 54.83 ± 16.57 ± 15.69 ± 14.95 ± 14.72 ± 15.30 ± 05.97 ± 05.24 ± 04.98 ± 05.53 ± 05.63 ± 03.56 ± n.a. 01.17 ± 28.24 ± 27.38 ± 26.98 ± 26.72 ± 25.29 ± 26.12 ± 02.90 ± 25.64 ± 19.45 ± 00.02 ± 01.95 ± 01.22 ± 00.51 ± n.a. 00.32 ±

17 70 232 174 29 5-Dec 27/43 120/112 133/41 20/09 65.94 ± 67.03 ± 66.25 ± 65.51 ± 65.90 ± 16.41 ± 17.01 ± 16.31 ± 15.68 ± 14.76 ± 05.85 ± 05.35 ± 05.61 ± 05.34 ± 05.78 ± 21.71 ± 29.08 ± 28.33 ± n.a. 26.42 ± 25.63 ± n.a. 25.46 ± 24.11 ± n.a. 27.29 ± 24.22 ± n.a. 02.07 ± 01.79 ± 00.88 ± 00.46 ± 00.44 ±

58 147 193 98 18 32/26 81/66 141/52 67/31 13/05 76.22 ± 74.96 ± 74.96 ± 73.49 ± 75.78 ± 15.90 ± 16.41 ± 16.30 ± 14.83 ± 13.00 ± 05.89 ± 05.12 ± 05.63 ± 05.37 ± 05.64 ± 23.32 ± 28.79 ± 27.92 ± n.a. 25.77 ± 25.00 ± n.a. 23.25 ± 23.61 ± n.a. 27.64 ± 20.22 ± n.a. 02.04 ± 01.67 ± 00.57 ± 00.44 ± 00.59 ±

08.88 07.62 05.62 07.51 08.72 02.35 02.54 02.52 02.98 04.75 01.15 00.59 01.09 01.17 01.18 03.58 00.79 01.36 01.30 02.29 02.29 02.78 03.82 03.86 02.83 07.50 10.73 00.15 00.45 00.67 00.67 00.60 00.59

01.33 02.27 01.30 02.90 03.02 02.29 02.42 02.19 02.82 03.48 01.27 00.69 01.23 01.37 01.55 21.71 28.33 01.30 01.92 01.84 02.78 04.67 03.30 02.83 06.73 10.16 00.15 00.40 00.42 00.54 00.48

03.03 02.43 02.83 02.97 03.04 02.55 02.13 02.53 02.91 05.51 00.66 00.56 01.16 01.26 00.92 03.27 00.73 01.33 02.26 02.39 03.99 04.21 08.15 11.67 00.50 00.77 00.86 00.72 00.57

02.95 02.96 02.80 02.19 03.06 02.34 02.70 02.52 03.26 04.41 01.16 00.56 00.96 01.16 02.14 03.59 00.81 01.38 02.47 02.22 05.52 06.52 07.19 09.31 00.40 00.59 00.70 00.54 00.61

Note: Data are presented as Mean ± Standard Deviation. AIG: Age-Independent Group; ADG: Age-Dependent Groups. N: Number; M/F: Male/Female; GDS: Geriatric Depression Scale; MMSE: Mini Mental State Examination; MoCA: Montreal Cognitive Assessment; Bjol: Benton Judgment of Line Orientation; MDS-UPDRS: Movement Disorder Society-Unified Parkinson’s Disease Rating Scale; H&Y: Hoehn and Yahr scale; MHIS: Modified Hachinski Ischemia Score. 3

Journal of Neuroscience Methods xxx (xxxx) xxx–xxx

G. Singh et al.

Here, μi and σi2 denote the ith image for class-wise mean and variance. Ni refers to number of images in ith class. We transformed this VIC image into input feature space for SOM by creating feature vectors. For each voxel, these feature vectors have information about x, y, z coordinates, a difference of intensity for the corresponding coordinates and the WAT score. See Eq.4.

likelihood of dementia due to vascular causes respectively. This information regarding subjects is provided in the supplementary information (Patient_Dataset.xlsx). 3. Note: HC subjects of age 51–60 years were obtained from PPMI and the rest from age 61–80 were obtained from ADNI

Xn = (i, j, k , IDij , Swat )

3.1. Image pre-processing

Here, Xn is nth sample from input feature space X, (i,j,k) are the 3D voxel coordinates with intensity IDij and WAT score Swat calculated using Eq. (4). Note: The range of the 3D voxel coordinates was normalized within [0, 1]. For 4th and 5th column, z-score based normalization was applied. SOM was used to create a discretized representation of every VIC image while still preserving the topographical relations and non-linear behavior present in the input data. Clustering in SOM is based on the principle of competitive learning. Based on numerical and topological similarity, high-dimensional input feature space are quantized and represented in the output space by best matching units (BMUS). Information about the input space vectors spanned by a BMUS neuron in the output space is called the Receptive Field (RF). SOM training is an iterative process with the following steps

Voxel-based morphometric analysis has been used for pre-processing and segmentation of brain MRIs into Grey Matter (GM) tissue images. To account for the large variation in brain anatomy amongst people (Skullerud, 1985), images were skull-stripped and co-registered by affine-transformation to ICBM template (Mazziotta et al., 1995). Thereafter, spatial normalization was performed using DARTEL template.4 Further, a spatially adaptive non- local means (SANLM) filter (Manjón et al., 2010) was also implemented for de-noising images while still preserving edges. This entire process was implemented by using’ Estimate and Write’ option in the VBM8 toolbox (Ashburner and Friston, 2000) for Statistical Parametric Mapping (SPM) (Penny et al., 2006) software v8.0. The size of the pre-processed brain image was 121 × 145 121 voxels.

1 Weight Initialization: At the start of training, random weights are assigned to neurons. This is done to initialize non-zero weights for the neurons. Custom weights may also be assigned to neurons. However, in our study, we choose random weight initialization. 2 Determine Best Matching Unit (BMUS): Each neuron in the output layer is presented to a random feature vector from the input layer feature space. The similarity of the output layer neurons with the input data instance is calculated using the squared Euclidean distance. This helps to determine a winning neuron that is also known as the Best Matching unit, see Eq. (5).

3.2. Feature extraction The main idea behind feature extraction is to identify biomarkers that can help to differentiate between given subject classes. However, there may exist several similar features corresponding to age-related changes between subject classes. Eliminating these common features would have two-fold benefits 1 Data quantization 2 Reduction of computational complexity

‖Xn −mc ‖ = min |x n−mi| For each comparison between subject classes, an image representing Voxel Intensity Changes (VIC) was created by subtracting the mean images of subjects for each class respectively, see Eqs. (1) and (2).

Here, Xn is nth sample from input feature space, mc is the reference vector closest to Xn and mi is the winning neuron

N

μi =

1 ∑ (Iki ) N k=1

3 Neighborhood determination: A unique property of SOM is that it takes into account topographical organization in the input data. To do this, after the winning neuron is determined, the weights of the neighboring neurons are updated as per a chosen kernel. Different kernels can be used to determine the neighbors of the winning neuron such as Bubble, Gaussian, and others. We have used a Gaussian kernel

Here, μi denotes the ith mean image of subject images in a class. N refers to the number of subjects and Iki refers to kth image in the ith class.

IDij = μi − μj Here, IDij is VIC image formed by subtracting mean images, μi and μj. Here,’ i’ and’ j’ are two different subject classes. Figure S1 shows schematic diagram for constructing VIC image. Now, all the features in the VIC image may not statistically useful to differentiate disease modalities. In our case, the number of subjects in disease classes was also unequal. To determine most discriminative features we have used score metric named Welch-Aspin Test (WAT), as reported earlier by (Singh and Lakshminarayanan, 2015) to be useful for statistical analysis of unbalanced classes. Increase in the numerical value of WAT scores directly corresponds to the statistical significance of the voxel’s discriminative ability. Class-wise images for mean and standard deviation for subjects were used to calculate WAT score for each pixel as follows

Swat =

hci (t ) = exp(−

Here, σt is the radius for neighborhood of the winning neuron at time t. dci2 is the lateral distance between the winning neuron and the neighborhood neuron. The size of this neighborhood radius is decreased with time for convergence. We have used exponential decay for the neighborhood radius.

σt = σ0exp(−

Ni

+

t ) τσ

The initial value of the neighborhood radius, σ0, can be manually chosen or be automatically selected by the SOM algorithm based on the chosen map size for the output layer of SOM.

μi − μ j σi2

dci2 ) 2 * σt2

σ 2j

4 Weight update and Learning: After the determination of the winning neuron and the neighborhood radius, the weights for the neurons in the output layer are updated based on lateral interaction between them using the following equation

Nj

4 Diffeomorphic anatomical registration through exponentiated lie algebra (DARTEL) template was derived from 550 subjects. For more information please refer to http:// brain-development.org/ixi-dataset/.

mi (t + 1) = mi (t ) + hci (t )[Xn (t )−mi (t )] 4

Journal of Neuroscience Methods xxx (xxxx) xxx–xxx

G. Singh et al.

4. Experimental design

Here, h ci (t) is the neighborhood function, t is the index of the iteration to map input space to output space. These steps from 2 to 4 are carried out iteratively, resulting in a low dimensional mapping of the input feature space. A different number of data points might be represented by BMUS at a region of interest (ROI). For normalization, the summation of intensity value at each ROI was divided by the number of the input space units clustered as BMUS, see Eq. (9).

ROIm =

The subject images were arranged into groups comprising of two subject classes at a time. For every group, we have made comparisons for Grey Matter (GM) brain tissue by creating Voxel Intensity Change (VIC) images, See Material and Methods. Further, the dataset was divided into two different ways viz. 1 Age-Independent Groups (AIG): Based on clinically identified subject classes, the entire cohort was divided into six groups viz. (i) AD vs. HC (ii) AD vs. MCI (iii) MCI vs. HC (iv) PD vs. HC (v) PD vs. SWEDD (vi) SWEDD vs. HC. These groups are named as AgeIndependent Groups (AIGs). 2 Age-Dependent Groups (ADG): AIGs were further divided into ADGs by considering an age range of 10 years resulting information of three age divisions for every AIG viz. 51–60, 61–70 and 71–80 years. A total of 18 Age-Dependent Groups (ADGs) were formed (6 binary classification groups × 3 age-divisions).

∑(i, j, k ) ∈ BMUS Im (i, j, k ) ∑(i, j, k ) ∈ BMUS NBMUS

Here, ROIm refers to normalized values for regions of interest selected by SOM and (i, j, k) are the 3D coordinates of each voxel corresponding to mth image. NBMUS refers to number of input space vectors clustered together as one best matching unit in the output space. These steps were carried out for each disease class. Thereafter, all the images for chosen classification group were projected onto the selected ROI space reducing every image to 300 data-points. Further, by using the information about 3D coordinates brain areas corresponding to identified ROIs were labeled by using Talairach Daemon (Lancaster et al., 2000).

AIGs were formed to identify important ROIs to aid differential diagnosis while comparing subject classes at large whereas, ADGs were formed to determine the trajectory for progression of NDs by including age as a factor. Table 1 shows clinical and demographic details of the subjects in these groups. While dividing subjects into AIGs, all the images obtained from data repositories for a subject class were considered at once, unlike ADGs wherein age was also included as a factor. VIC images, created for ADGs represent age-related snapshots whereas, for AIGs, it provides an overall estimate of brain areas involved in the disease for subject classes under consideration. VIC images for all AIGs and ADGs were used to identify disease-related biomarkers that could be used to establish the differential diagnosis. As shown in, Brain areas detected using the VIC image for ADGs are shown with specific color as follows (a) 51–60 ( ) (b) 61–70 ( ) (c) 71–80 ( ) whereas, Biomarkers detected for AIGs are shown under the heading “51–80 ( )”.

3.3. Feature classification In this study, we have used Least Square Support Vector Machine (LSSVM) for subject classification. LSSVM is a supervised machine learning algorithm for classification. It is a least square variant of the traditional SVM approach. Here, equality constraint is applied to the decision variable resulting in the possibility of the implication of the Kuhn-tucker boundary conditions to the Lagrange multiplier. This makes it possible to solve a set of linear equations unlike in the case of SVM where a quadratic problem needs to be solved for optimization. The optimization is aimed at determining a maximum-margin hyperplane between the input space classes. To allow mapping of non-linear separating boundary between the classes, kernel functions may be used. These kernel functions project the input vectors to a high-dimensional feature space to determine the separating hyperplane. In this study, we have used the radial basis function (RBF) kernel function.

K (x , xk ) = exp(

−x‖2

−‖xk σ2

4.1. Selection of optimal number of features Data dimensionality reduction is important as rarely all of the features are critical for classification. Further, owing to the similarity between unaffected areas of the brain, removal of redundant features not only speeds up the learning process but also helps in improving the classification accuracy (Chu et al., 2012). To determine an optimal choice, we performed individual-level classification by varying the number of neurons in SOM output space from 1 to 10,000. Results for all the classes were averaged to obtain a common choice of 300 neurons to be selected by SOM (see Fig. 1). For each comparison we calculated 6 metrics namely Accuracy (ACC), True Positive Ratio (TPR), True Negative Ratio (TNR), Positive Predictive Value (PPV), Negative Predictive Value (NPV), Mathews Correlation Coefficient (MCC). Average of the former 5 and MCC were used to adjudge the quality of the binary classification and determine the optimal number of features to be used. These metrics have been plotted in Fig. 1. Ten-fold cross-validation was performed for each experiment. Table S7 shows detailed mean ± standard deviation of each experiment. The metrics used are defined as follows

)

Here, x refers to the input space vector, xk is the kth vector for hyperplane boundary, σ is a RBF kernel parameter that is optimized during model training. The classifier model is iteratively trained to fit the following relationship between the input space features, x, and the dependent output variable, y, minimizing the cost function, J, at the same time.

y (x ) = ωT K (x , xk ) + b + ek

J=

1 T ω ω+γ 2

N

∑ ek2 k=1

Here, ω is the regression weight, ek2 is the squared regression error, γ is regularization parameter. All the images in a disease class were combined together to obtain data matrices for every comparison, consisting of n X 300 data points, where’ n’ represents the number of subjects for each class. Every dataset was randomly divided into 10 parts. 80% of these images were modelled using SOM to identify most relevant ROIs useful for separating the two classes under consideration. This formed the training set for LSSVM.5 Rest of the images were used as the test set and the classifier performance on this set has been reported. 5

LSSVM toolbox was obtained from http://www.esat.kuleuven.be/sista/lssvmlab/.

5

ACC =

TP + TN TP + TN + FP + FN

TPR =

TP TP + FN

TNR =

TN TN + FP

Journal of Neuroscience Methods xxx (xxxx) xxx–xxx

G. Singh et al.

Fig. 1. Classification metrics obtained upon varying number of features presented to LS-SVM. We have made a tissue-wise comparison for AD vs. HC, AD vs. MCI, AD vs. PD, MCI vs. HC, PD vs. HC, PD vs. SWEDD and SWEDD vs. HC. Optimal choice for number of features to be used in the final model was based on obtaining max {average (Accuracy, True Positive Rate (TPR), True Negative Rate (TNR), Positive Predictive Value (PPV) and Negative Predictive Value (NPV))}, see Table S7. Additionally, we have also plotted Mathews Correlation Coefficient (MCC) for assessment of quality of binary classification. In (A) MCC and the average of the 5 metrics and in (B) feature wise plot for the 5 metrics considered in this study have been plotted.

PPV =

TP TP + FP

NPV =

TN TN + FN

MCC =

(TP * TN )− (FP * FN ) (TP + FP )(TP + FN )(TN + FP )(TN + FN )

In other cases, for areas identified in any ADG or AIG, Z-scores of the other three ADGs and AIG for the same comparison have been reported. Color-coding for a particular brain area indicate the age range during which the brain area was identified to be involved in the pathology of diseases being compared. Also, see supplementary videos for disease progression. 5.1. Biomarkers for differential diagnosis of AD and related disorders

Here, for each binary comparison, considering one class as positive and the other one as negative TP, TN, FP, and FN are defined as True Positives, True Negatives, False Positives, False Negatives respectively.

For AD Pathology, We identified the involvement of Anterior lobe, Basal Ganglia, Cerebellum, Diencephalon, Limbic Lobe, Mesencephalon and Occipital Lobe. Specific areas, their location, and Z-scores have been tabulated in Table S1. In accordance with the previous findings, Brodmann Area 23 (BD23), part of Posterior Cingulate Cortex (PCC), was one of the earliest areas found to be affected in the AD (Zhou et al., 2008), See Fig. 3. Similarly, Retrosplenial Cortex (RSC) is known to be compromised in NDs, like AD, that involves memory impairment (Vann et al., 2009). Reduction in metabolic activity in areas belonging to RSC, namely Brodmann Area 29 (BD29) and 30 (BD30) has been reported. As a part of the Papez circuit, RSC receives projections from the Anterior Thalamic nuclei (ATN). ATN has also been reported to be involved in the AD pathogenesis (De Jong et al., 2008). At large, Thalamus is involved in sensory and motor responses. Specifically, it’s Anterior (ATN), Medial-Dorsal (MDN), Intralaminar and Midline nuclei (MN) are known to be involved in regulation of memory functions (Van Der Werf et al., 2000). Jong et. al. reported a significant reduction in the volumes of Putamen and Thalamus in patients diagnosed with probable AD. This decrease in volume was also found to be linearly correlated with

5. Results & discussion Fig. 2 shows the ecosystem of imaging biomarkers for NDs. Corresponding to each classification group, brain areas have been highlighted in a color-coded manner where the color of the box represents the age range at which the brain area was detected to be involved in the subject class(es) under consideration (See Section 3.1). Pairs of arrows inside every box represent the change i.e. hyperintensity (↑) and hypointensity (↓), in left and right hemisphere of the brain area under consideration. Hyperintensity and hypoactivation represent hyperactivation and degeneration of the brain area under comparison respectively. In total, four pairs of arrows represent the state of activity for the brain area at 51–60, 61–70, 71–80 years and 51–80 years respectively. Table 2 shows brain areas that were identified to be most important for a differential diagnosis by using our automated algorithm. In many cases, brain areas identified for different ADGs overlapped. 6

Journal of Neuroscience Methods xxx (xxxx) xxx–xxx

G. Singh et al.

Fig. 2. Ecosystem of brain imaging biomarker for neurodegenerative diseases. Brain imaging biomarkers areas were detected using the VIC image for ADGs and AIGs. Specific color used for age-groups is as follows (a) 51–60 ( ) (b) 61–70 ( ) (c) 71–80 ( ) (d) 51–80 ( ). The disease trajectories for all the diseases compared in this study have been shown.

In Basal Ganglia, we observed changes in the Caudate Body (CB) and Declive (DV) in accordance with (Ryan et al., 2013). The perforant pathway in the brain connects the Entorhinal Cortex with the Hippocampal formation. (Seifan et al., 2015) has postulated that tauopathic burden within specific layers of the Hippocampus could be due to projections of the lateral Entorhinal Cortex to the outer molecular layer of the Dentate Gyrus (DT). In line with this, we observed atrophy in areas Fastigium (FT) and Nodule of Vermis (NoV). Copenhaver et al. reported insignificant changes for Mammillary Bodies (MB) in preclinical stages of the dementing process, but clear changes in the AD (Copenhaver et al., 2006). It was interesting to note that even in our study the involvement of MB was detected in later part of the disease i.e 71–80 years. Mild cognitive impairment is an intermediate stage between normal cognitive decline and AD. MCI patients are at a greater risk of developing AD. As expected, for the brain areas affected in early stages, we found many similarities between AD and MCI subjects such as BD23,

impaired global cognitive performance (De Jong et al., 2008). Atrophy in other parts of Thalamus such as MDN (De Jong et al., 2008; Kenny et al., 2012; Li et al., 2014; Ryan et al., 2013; Vann et al., 2009), Ventral-Lateral (VLN) (De Jong et al., 2008), Pulvinar (PV) (Stratmann et al., 2016) and MN (Engelhardt and Laks, 2008) has been reported. We also observed similar changes in these brain areas. In the Limbic Lobe, we detected the atrophy of Brodmann Area 33 (BD33), part of Anterior Cingulate Cortex (ACC). In support of this, Bailly et.al while comparing glucose metabolism and atrophy, had observed significantly smaller 18FFDG uptake in ACC in AD patients than in HC (Bailly et al., 2015). Using voxel-based morphometry, Kinkingnéhun et. al. conducted a 3-year follow-up for mild AD dichotomized into slow decliners (SLD) or fast decliners (FD) groups on the basis of their decline in Mini-Mental State Examination score over time (Kinkingnéhun et al., 2008). They reported significant atrophy in medial occipitoparietal areas for FD than SLD patients. We also observed atrophy of Cerebral Lingual (CL) and Culmen Of Vermis (CoV). 7

Journal of Neuroscience Methods xxx (xxxx) xxx–xxx

G. Singh et al.

Fig. 3. Common brain areas affected during AD and PD. For Differential diagnosis, it is important to identify biomarkers that can be useful to differentiate between diseases. Thus, similar brain changes must be identified. These changes could be due to normal ageing process. (A) We found RN, CoV, MDN and PV to be affected during the earliest stage of AD related disease. (B) For parkinson related disorders, we found PV and MDN to be affected in all the related disorders. These brain areas might not be useful to differentiate AD and PD related disorders. Hence, changes in the brain areas shown in (A) and (B) might not indicate AD/PD disease progression respectively.

for RSC was positively correlated with motor and cognitive decline in non-demented PD patients (Nagano-Saito et al., 2004). We found BD29 (part of RSC) to be one of the earliest areas affected in PD along with CoV, MDN, PV and Ventral Anterior Nucleus (VAN). Hsu et al. also reported the involvement of CoV and CL in PD by applying Independent Component Analysis (ICA) to assess the difference of rCBF between PD and HC subjects (Hsu et al., 2003). While investigating the microstructural integrity of thalamic regions in de novo PD patients relative to HC, Planetta et al. found significant reduction in functional anisotropy values in the fibers projecting from the ATN, VAN, and MDN in early stages of PD group, but not in the Lateral Posterior Nucleus (LPN), Ventral Posterior Medial Nucleus (VPMN) and PV (Planetta et al., 2013), See Fig. 3. We also found that LPN and VPMN were found to be involved only during 71–80 years. Contrastingly, in our study we found PV nuclei to be amongst the earliest areas affected in PD (Pifl et al., 2012) (See Table S4). SWEDDs are a group of subjects that exhibit clinical symptoms similar to PD, even though the underlying pathophysiological process differs from PD. Thus, SWEDDs are often misdiagnosed as PD (Bajaj

MDN, PV, RN (See Table S2). Other areas, such as CB, MN, and FT were not detected while comparing MCI and HC. However, on comparing MCI and AD subjects, these areas were again found to be involved (See Table S3). This may be indicative of the lower involvement of these areas in subjects suffering MCI as compared to the AD. Another interesting observation was that RN was found to be involved in both AD/ MCI vs. HC, however, on comparing AD vs. MCI, RN was found to be involved in very late stages of the disease. This could be due to the fact that in our methodology brain areas are identified using VIC image obtained by subtracting average images of the two groups. Since RN was involved in early stages for both AD and MCI, it was not detected until a further deterioration in late stages of the AD. 5.2. Biomarkers for differential diagnosis of PD and related disorders In most of the PD patients, along with established clinical symptoms such as tremors, bradykinesia, change in speech and other cognitive and motor dysfunction is also observed. Recently, Nagano-Saito et al. also reported that Raven’s Coloured Progressive Matrices (RCPM) score 8

Journal of Neuroscience Methods xxx (xxxx) xxx–xxx

G. Singh et al.

Table 2 Biomarkers detected for differential diagnosis of neurodegenerative diseases considered in this study.

Brain imaging biomarkers areas were detected using the VIC image for ADGs and AIGs. Specific color used for age-groups is as follows (a) 51–60 ( ) (b) 61–70 ( ) (c) 71–80 ( ) (d) 51–80 ( ). Pairs of arrows inside every box represent the change i.e. Hyperintensity (↑) and Hypointensity (↓), in left and right hemisphere of the brain area under consideration. Hyperintensity and hypoactivation represent hyperactivation and degeneration of the brain area under comparison respectively. In total, four pairs of arrows represent the state of activity for the brain area at 51–60, 61–70, 71–80 years and 51–80 years respectively.

et al., 2010). On comparing SWEDDs with HC, we only found BD23, MDN, and PV to be affected in the early stages (Singh and Lakshminarayanan, 2015) unlike PD vs HC where additionally RSC, CoV, and VAN were also found to be affected. In SWEDDs, RSC was found to be affected in the later part of the disease. While RN was affected by both diseases separately, as expected, SN was found to be affected only in PD, See Table S5 and S6.

we have achieved 94.05 ± 0.05, 92.55 ± 0.07, 79.91 ± 0.16 for average accuracy, sensitivity and specificity respectively. This establishes that the methodology is not disease specific but can perform well on multiple diseases. This is owing to the fact that we did not initially restrict our choice of ROIs but performed a whole-brain analysis. Thus, this helps to remove the individual specific bias and provides a more holistic understanding of the underlying pathology of diseases.

5.3. Individual-level subject classification

6. Conclusion

The potential of our methodology is not limited to detecting biomarkers but has also shown promising results for classifying subjects into probable subject classes. An overall classification accuracy of > 90% has been achieved in distinguishing between the subject classes considered in this study. Additionally, we present here a comparison between classification accuracy achieved by other studies investigating the potential of machine learning based approach for clinical decision support system (see Table 3). In most of the comparisons, the classification accuracy of > 90% has been achieved. Specifically, we obtained highest classification accuracy of 94.29 ± 0.08 and 95.37 ± 0.02 for distinguishing AD and PD from HC subjects respectively. For distinguishing disease classes with the similar clinical presentation, our algorithm achieved 92.46 ± 0.03 and 96.04 ± 0.03 for distinguishing the AD from MCI and PD from SWEDD subjects respectively. It is interesting to note that for most of the comparison, our study had the largest cohort of subjects. Also, the methodology performed well for both the AD, PD and their related counterparts. Specifically for AD and related counterparts, we have achieved 92.01 ± 0.05, 86.51 ± 0.10, 90.03 ± 0.07 whereas for PD and related counterparts

Unlike most of the studies conducted in the past that focused only on one major disease at once i.e. AD or PD, in this work we have compared diseased subjects belonging to both of these prevalent diseases with healthy controls or other disease subjects. Specifically, AD and PD have been compared with HC and two other diseases that exhibit similar signs and symptoms, i.e., MCI and SWEDD. In accordance with previous findings in the literature several imaging biomarkers were detected to be of importance for differential diagnosis. Relevant brain areas were automatically detected without the need of a priori information. This highlights the benefits of the proposed framework that is non-specific to any particular disease but can be easily extended to other NDs as well. Overall several important conclusions can be drawn, firstly, brain areas detected in the earliest stage ADGs, i.e. 51–60 years, may serve as useful biomarkers for early-stage disease diagnosis. Secondly, these biomarkers can be used to estimate disease trajectory. Thirdly, unlike manual selection of ROIs, we have used an unbiased approach wherein brain areas were automatically detected without any a priori information. Fourthly, as per our knowledge, this is the first time when a machine learning method has been combined with the MRI-based 9

Journal of Neuroscience Methods xxx (xxxx) xxx–xxx

G. Singh et al.

Table 3 Comparison of studies investigating potential of machine learning based computer-assisted decision support system. Subjects

N

Accuracy

TPR

TNR

AD vs. HC

128/262 128/262 75/75 58/28 182/226 32/18 190/190 51/52 128/447 128/447 21/15 447/262 447/262 15/20 302/189 79/204 15/15 15/15 23/25 67/35 24/18 99/52 408/262 408/262 518/245 22/21 28/28 408/71 408/71 518/68 71/262 71/262 68/245

94.29 ± 0.08 91.54 ± 0.05 92 84 90.5 90 89.3 93.2 85.43 ± 0.08 92.46 ± 0.03 87 95.24 ± 0.05 93.14 ± 0.02 95 92 71.09 90 100 83 84 97.62 76.4 92.63 ± 0.06 95.37 ± 0.02 93.25 ± 0.46 39.53 85.8 94.63 ± 0.05 96.04 ± 0.03 99.86 ± 0.10 92.65 ± 0.08 93.03 ± 0.05 100.00 ± 0.00

88.11 ± 0.18 95.77 ± 0.05 – 86 85 97 88 93 81.33 ± 0.14 73.85 ± 0.10 85 93.45 ± 0.10 86.54 ± 0.04 93 89 51.96 – – 83 89 96 81.8 94.18 ± 0.05 91.15 ± 0.06 97.12 45.45 86 100.00 ± 0.00 100.00 ± 0.00 100 70.00 ± 0.33 100.00 ± 0.00 100

94.78 83.08 – 82 94.8 78 90 93.3 79.65 97.95 80 92.54 97.05 90 80 78.4 – – 84 85 100 66 81.80 98.05 85.03 33.33 86 64.44 72.86 98.81 99.17 67.14 100

AD vs. MCI

MCI vs. HC

PD vs. HC

PD vs. SWEDD

SWEDD vs. HC

Authors ± 0.08 ± 0.11

± 0.16 ± 0.02 ± 0.06 ± 0.03

± 0.16 ± 0.02

± 0.32 ± 0.23 ± 0.03 ± 0.22

AIG ADG Duchesne et al. (2008) Ferrarini et al., (2008) Nho et al. (2010) Plant et al. (2010) Vemuri et al. (2008) Zhang et al. (2011) AIG ADG Chen et al. (2011) AIG ADG Chen et al. (2011) Chincarini et al. (2011) Cui et al. (2012) Davatzikos et al. (2008) Fan et al. (2008) Gerardin et al. (2009) Haller et al. (2010) Plant et al. (2010) Zhang et al. (2011) AIG ADG Singh and Lakshminarayanan (2015) Duchesne et al. (2009) Salvatore et al. (2014) AIG ADG Singh and Lakshminarayanan (2015) AIG ADG Singh and Lakshminarayanan (2015)

Note: Data for this study has been presented as Mean ± Standard Deviation. AIG: Age-Independent Group; ADG: Age-Dependent Groups. N: Number; TPR: True positive rate also known as Sensitivity; TNR: True negative rate also known as Specificity; PPV: Positive predictive value; NPV: Negative predictive value.

investigators within the ADNI and PPMI only provided the data but they did not contribute towards analysis or writing of this report.

approach in a single study the etiology of the AD, PD, and their clinically similar counterparts. Lastly, the adaptability of this methodology to detect relevant brain areas, not only for the disease classes described in this study but for other NDs, makes it an ideal candidate for clinical translation.

Appendix A. Supplementary data Supplementary material related to this article can be found, in the online version, at doi:https://doi.org/10.1016/j.jneumeth.2018.05. 009.

Author contributions G.S. and L.S. conceived the study and the methodology; G.S. prepared the data sets, designed & implemented the methodology, and prepared the first draft of the manuscript; E.C.L. offered inputs from neurologist’s perspective and helped in the interpretation of obtained results; L.S. edited the manuscript multiple times and worked with G.S. to produce the final version of the manuscript.

References Ashburner, J., Friston, K.J., 2000. Voxel-based morphometry - the methods. Neuroimage 11, 805–821. http://dx.doi.org/10.1006/nimg.2000.0582. Bailly, M., Destrieux, C., Hommet, C., Mondon, K., Cottier, J.-P., Beaufils, E., Vierron, E., Vercouillie, J., Ibazizene, M., Ziane, Voisin, T., Payoux, P., Barre, L., Camus, V., Guilloteau, D., Ribeiro, M.J., 2015. Precuneus and cingulate cortex atrophy and hypometabolism in patients with Alzheimer's disease and mild cognitive impairment: MRI and 18F-FDG PET quantitative analysis using FreeSurfer. Biomed. Res. Int. 2015 (2015), e583931. http://dx.doi.org/10.1155/2015/583931. Bajaj, N.P.S., Gontu, V., Birchall, J., Patterson, J., Grosset, D.G., Lees, A.J., 2010. Accuracy of clinical diagnosis in tremulous parkinsonian patients: a blinded video study. J. Neurol. Neurosurg. Psychiatry 81, 1223–1228. http://dx.doi.org/10.1136/ jnnp.2009.193391. Chen, G., Ward, B.D., Xie, C., Li, W., Wu, Z., Jones, J.L., Franczak, M., Antuono, P., Li, S.J., 2011. Classification of Alzheimer disease, mild cognitive impairment, and normal cognitive status with large-scale network analysis based on resting-state functional MR imaging. Radiology 259, 213–221. http://dx.doi.org/10.1148/radiol. 10100734. Chincarini, A., Bosco, P., Calvini, P., Gemme, G., Esposito, M., Olivieri, C., Rei, L., Squarcia, S., Rodriguez, G., Bellotti, R., Cerello, P., De Mitri, I., Retico, A., Nobili, F., 2011. Local MRI analysis approach in the diagnosis of early and prodromal Alzheimer’s disease. Neuroimage 58, 469–480. http://dx.doi.org/10.1016/j. neuroimage.2011.05.083. Chu, C., Hsu, A.L., Chou, K.H., Bandettini, P., Lin, C., 2012. Does feature selection

Funding G.S. would like to acknowledge the support of the National University of Singapore by providing NUS Research Scholarship towards the fulfillment of Doctor of Philosophy. Acknowledgments Data collection and sharing for this study was funded by the Alzheimer’s disease Neuroimaging Initiative http://adni.loni.usc.edu/ and the Parkinson’s Progression Markers Initiative http://www.ppmiinfo.org/ primarily supported by National Institute on Aging, the National Institute of Biomedical Imaging & Bioengineering and by the Michael J. Fox Foundation for Parkinson’s Research respectively. The 10

Journal of Neuroscience Methods xxx (xxxx) xxx–xxx

G. Singh et al.

Kochunov, P.V., Nickerson, D., Mikiten, S.A., Fox, P.T., 2000. Automated Talairach Atlas labels for functional brain mapping. Hum. Brain Mapp. 10, 120–131. http://dx. doi.org/10.1002/1097-0193(200007)10:3<120::AID-HBM30>3.0.CO;2-8. Li, W., Antuono, P.G., Xie, C., Chen, G., Jones, J.L., Ward, B.D., Singh, S.P., Franczak, M.B., Goveas, J.S., Li, S.J., 2014. Aberrant functional connectivity in Papez circuit correlates with memory performance in cognitively intact middle-aged APOE4 carriers. Cortex 57 167–176. http://dx.doi.org/10.1016/j.cortex.2014.04.006. Manjón, J.V., Coupé, P., Martí-Bonmatí, L., Collins, D.L., Robles, M., 2010. Adaptive nonlocal means denoising of MR images with spatially varying noise levels. J. Magn. Reson. Imaging 31, 192–203. http://dx.doi.org/10.1002/jmri.22003. Marek, K., Jennings, D., Lasch, S., Siderowf, A., Tanner, C., Simuni, T., Coffey, C., Kieburtz, K., Flagg, E., Chowdhury, S., Poewe, W., Mollenhauer, B., Sherer, T., Frasier, M., Meunier, C., Rudolph, A., Casaceli, C., Seibyl, J., Mendick, S., Schuff, N., Zhang, Y., Toga, A., Crawford, K., Ansbach, A., de Blasio, P., Piovella, M., Trojanowski, J., Shaw, L., Singleton, A., Hawkins, K., Eberling, J., Russell, D., Leary, L., Factor, S., Sommerfeld, B., Hogarth, P., Pighetti, E., Williams, K., Standaert, D., Guthrie, S., Hauser, R., Delgado, H., Jankovic, J., Hunter, C., Stern, M., Tran, B., Leverenz, J., Baca, M., Frank, S., Thomas, C.A., Richard, I., Deeley, C., Rees, L., Sprenger, F., Lang, E., Shill, H., Obradov, S., Fernandez, H., Winters, A., Berg, D., Gauss, K., Galasko, D., Fontaine, D., Mari, Z., Gerstenhaber, M., Brooks, D., Malloy, S., Barone, P., Longo, K., Comery, T., Ravina, B., Grachev, I., Gallagher, K., Collins, M., Widnell, K.L., Ostrowizki, S., Fontoura, P., La-Roche, F.H., Ho, T., Luthman, J., van der Brug, M., Reith, A.D., Taylor, P., 2011. The Parkinson progression marker initiative (PPMI). Prog. Neurobiol. 95, 629–635. http://dx.doi.org/10.1016/j. pneurobio.2011.09.005. Mazziotta, J.C., Toga, A.W., Evans, A., Fox, P., Lancaster, J., 1995. A Probabilistic Atlas of the human brain: theory and rationale for its development. Neuroimage 2, 89–101. http://dx.doi.org/10.1006/nimg.1995.1012. Nagano-Saito, A., Kato, T., Arahata, Y., Washimi, Y., Nakamura, A., Abe, Y., Yamada, T., Iwai, K., Hatano, K., Kawasumi, Y., Kachi, T., Dagher, A., Ito, K., 2004. Cognitive- and motor-related regions in Parkinson's disease: FDOPA and FDG PET studies. Neuroimage 22, 553–561. http://dx.doi.org/10.1016/j.neuroimage.2004.01.030. Nho, K., Shen, L., Kim, S., Risacher, S.L., West, J.D., Foroud, T., Jack, C.R., Weiner, M.W., Saykin, A.J., 2010. Automatic prediction of conversion from mild cognitive impairment to probable Alzheimer’s disease using structural magnetic resonance imaging. AMIA Annual Symposium Proceedings. pp. 542–546. Penny, W.D., Friston, K.J., Ashburner, J., Kiebel, S.J., Nichols, T., 2006. Statistical Parametric Mapping: The Analysis of Funtional Brain Images. Elsevier/Academic Press. Pifl, C., Kish, S.J., Hornykiewicz, O., 2012. Thalamic noradrenaline in Parkinson's disease: deficits suggest role in motor and non-motor symptoms. Mov. Disord. 27, 1618–1624. http://dx.doi.org/10.1002/mds.25109. Planetta, P.J., Schulze, E.T., Geary, E.K., Corcos, D.M., Goldman, J.G., Little, D.M., Vaillancourt, D.E., 2013. Thalamic projection fiber integrity in de novo Parkinson disease. Am. J. Neuroradiol. 34, 74–79. http://dx.doi.org/10.3174/ajnr.A3178. Plant, C., Teipel, S.J., Oswald, A., Böhm, C., Meindl, T., Mourao-Miranda, J., Bokde, A.W., Hampel, H., Ewers, M., 2010. Automated detection of brain atrophy patterns based on MRI for the prediction of Alzheimer’s disease. Neuroimage 50, 162–174. http:// dx.doi.org/10.1016/j.neuroimage.2009.11.046. Prince, M., Prina, M., Guerchet, M., 2013. Journey of Caring: An Analysis of Long-Term Care for Dementia. WHO Press, Alzheimer’s Disease International, Geneva, Switzerland. Ryan, N.S., Keihaninejad, S., Shakespeare, T.J., Lehmann, M., Crutch, S.J., Malone, I.B., Thornton, J.S., Mancini, L., Hyare, H., Yousry, T., Ridgway, G.R., Zhang, H., Modat, M., Alexander, D.C., Rossor, M.N., Ourselin, S., Fox, N.C., 2013. Magnetic resonance imaging evidence for presymptomatic change in thalamus and caudate in familial Alzheimer’s disease. Brain 136, 1399–1414. http://dx.doi.org/10.1093/brain/ awt065. Salvatore, C., Cerasa, A., Castiglioni, I., Gallivanone, F., Augimeri, A., Lopez, M., Arabia, G., Morelli, M., Gilardi, M.C., 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. http://dx.doi.org/10.1016/j.jneumeth. 2013.11.016. Seifan, A., Marder, K.S., Mez, J., Noble, J.M., Cortes, E.P., Vonsattel, J.P., Honig, L.S., 2015. Hippocampal laminar distribution of tau relates to Alzheimer’s disease and age of onset. J. Alzheimer’s Dis. 43, 315–324. http://dx.doi.org/10.3233/JAD-140279. Singh, G., Lakshminarayanan, S., 2015. Unsupervised learning based feature extraction for differential diagnosis of neurodegenerative diseases: a case study on early-stage diagnosis of Parkinson disease. J. Neurosci. Methods 256, 30–40. http://dx.doi.org/ 10.1016/j.jneumeth.2015.08.011. Skullerud, K., 1985. Variations in the size of the human brain. Influence of age, sex, body length, body mass index, alcoholism, Alzheimer changes, and cerebral atherosclerosis. Acta Neurol. Scand. Suppl. 102, 1–94. Stratmann, K., Heinsen, H., Korf, H.W., Del Turco, D., Ghebremedhin, E., Seidel, K., Bouzrou, M., Grinberg, L.T., Bohl, J., Wharton, S.B., Den Dunnen, W., Rüb, U., 2016. Precortical phase of Alzheimer’s Disease (AD)-Related tau cytoskeletal pathology. Brain Pathol. 26, 371–386. http://dx.doi.org/10.1111/bpa.12289. United Nations, 2013. World Population Ageing 2013. Van Der Werf, Y.D., Witter, M.P., Uylings, H.B.M., Jolles, J., 2000. Neuropsychology of infarctions in the thalamus: a review. Neuropsychologia 38, 613–627. http://dx.doi. org/10.1016/S0028-3932(99)00104-9. Vann, S.D., Aggleton, J.P., Maguire, E.A., 2009. What does the retrosplenial cortex do? Nat. Rev. Neurosci. 10, 792–802. http://dx.doi.org/10.1038/nrn2733. Vemuri, P., Gunter, J.L., Senjem, M.L., Whitwell, J.L., Kantarci, K., Knopman, D.S., Boeve, B.F., Petersen, R.C., Jack, C.R., 2008. Alzheimer's disease diagnosis in individual subjects using structural MR images: validation studies. Neuroimage 39, 1186–1197.

improve classification accuracy? Impact of sample size and feature selection on classification using anatomical magnetic resonance images. Neuroimage 60, 59–70. http://dx.doi.org/10.1016/j.neuroimage.2011.11.066. Copenhaver, B.R., Rabin, L.A., Saykin, A.J., Roth, R.M., Wishart, H.A., Flashman, L.A., Santulli, R.B., McHugh, T.L., Mamourian, A.C., 2006. The fornix and mammillary bodies in older adults with Alzheimer's disease, mild cognitive impairment, and cognitive complaints: a volumetric MRI study. Psychiatry Res. - Neuroimaging 147, 93–103. http://dx.doi.org/10.1016/j.pscychresns.2006.01.015. Cui, Y., Wen, W., Lipnicki, D.M., Beg, M.F., Jin, J.S., Luo, S., Zhu, W., Kochan, N.A., Reppermund, S., Zhuang, L., Raamana, P.R., Liu, T., Trollor, J.N., Wang, L., Brodaty, H., Sachdev, P.S., 2012. Automated detection of amnestic mild cognitive impairment in community-dwelling elderly adults: a combined spatial atrophy and white matter alteration approach. Neuroimage 59, 1209–1217. http://dx.doi.org/10.1016/j. neuroimage.2011.08.013. Davatzikos, C., Fan, Y., Wu, X., Shen, D., Resnick, S.M., 2008. Detection of prodromal Alzheimer’s disease via pattern classification of magnetic resonance imaging. Neurobiol. Aging 29, 514–523. http://dx.doi.org/10.1016/j.neurobiolaging.2006.11. 010. De Jong, L.W., Van Der Hiele, K., Veer, I.M., Houwing, J.J., Westendorp, R.G.J., Bollen, E.L.E.M., De Bruin, P.W., Middelkoop, H.A.M., Van Buchem, M.A., Van Der Grond, J., 2008. Strongly reduced volumes of putamen and thalamus in Alzheimer's disease: an MRI study. Brain 131, 3277–3285. http://dx.doi.org/10.1093/brain/awn278. Dorsey, E.R., Constantinescu, R., Thompson, J.P., Biglan, K.M., Holloway, R.G., Kieburtz, K., Marshall, F.J., Ravina, B.M., Schifitto, G., Siderowf, A., Tanner, C.M., 2007. Projected number of people with Parkinson disease in the most populous nations, 2005 through 2030. Neurology 68, 384–386. http://dx.doi.org/10.1212/01.wnl. 0000247740.47667.03. Duchesne, S., Caroli, A., Geroldi, C., Barillot, C., Frisoni, G.B., Collins, D.L., 2008. MRIbased automated computer classification of probable AD versus normal controls. IEEE Trans. Med. Imaging 27, 509–520. http://dx.doi.org/10.1109/TMI.2007.908685. Duchesne, S., Rolland, Y., Vérin, M., 2009. Automated computer differential classification in parkinsonian syndromes via pattern analysis on MRI. Acad. Radiol. 16, 61–70. http://dx.doi.org/10.1016/j.acra.2008.05.024. Engelhardt, E., Laks, J., 2008. Alzheimer disease neuropathology: understanding autonomic dysfunction. Dement. Neuropsychol. 2, 183–191. http://dx.doi.org/10.1590/ S1980-57642009DN20300004. Fan, Y., Resnick, S.M., Wu, X., Davatzikos, C., 2008. Structural and functional biomarkers of prodromal Alzheimer's disease: a high-dimensional pattern classification study. Neuroimage 41, 277–285. http://dx.doi.org/10.1016/j.neuroimage.2008.02.043. Ferrarini, L., Palm, W.M., Olofsen, H., Van Der Landen, R., Van Buchem, M.A., Reiber, J.H.C., Admiraal-Behloul, F., 2008. Ventricular shape biomarkers for Alzheimer’s disease in clinical MR images. Magn. Reson. Med. 59, 260–267. http://dx.doi.org/10. 1002/mrm.21471. Gerardin, E., Chételat, G., Chupin, M., Cuingnet, R., Desgranges, B., Kim, H.S., Niethammer, M., Dubois, B., Lehéricy, S., Garnero, L., Eustache, F., Colliot, O., 2009. Multidimensional classification of hippocampal shape features discriminates Alzheimer’s disease and mild cognitive impairment from normal aging. Neuroimage 47, 1476–1486. http://dx.doi.org/10.1016/j.neuroimage.2009.05.036. Graña, M., Termenon, M., Savio, A., Gonzalez-Pinto, A., Echeveste, J., Pérez, J.M., Besga, A., 2011. Computer aided diagnosis system for Alzheimer disease using brain diffusion tensor imaging features selected by Pearson’s correlation. Neurosci. Lett. 502, 225–229. http://dx.doi.org/10.1016/j.neulet.2011.07.049. Haller, S., Bartsch, A., Nguyen, D., Rodriguez, C., Emch, J., Gold, G., Lovblad, K.O., Giannakopoulos, P., 2010. Cerebral microhemorrhage and iron deposition in mild cognitive impairment: susceptibility-weighted MR imaging assessment. Radiology 257, 764–773. http://dx.doi.org/10.1148/radiol.10100612. Hsu, J.-L., Duann, J.-R., Wang, H.-C., Jung, T.-P., 2003. Assessing rCBF changes in Parkinson’s disease using independent component analysis, Nara, Japan. Prodceedings of ICA 2003 873–878. Jack, C.R., Bernstein, M.A., Fox, N.C., Thompson, P., Alexander, G., Harvey, D., Borowski, B., Britson, P.J., Whitwell, J.L., Ward, C., Dale, A.M., Felmlee, J.P., Gunter, J.L., Hill, D.L.G., Killiany, R., Schuff, N., Fox-Bosetti, S., Lin, C., Studholme, C., DeCarli, C.S., Krueger, G., Ward, H.A., Metzger, G.J., Scott, K.T., Mallozzi, R., Blezek, D., Levy, J., Debbins, J.P., Fleisher, A.S., Albert, M., Green, R., Bartzokis, G., Glover, G., Mugler, J., Weiner, M.W., 2008. The Alzheimer's disease neuroimaging initiative (ADNI): MRI methods. J. Magn. Reson. Imaging 27, 685–691. http://dx.doi.org/10.1002/jmri. 21049. Jack, C.R., Wiste, H.J., Vemuri, P., Weigand, S.D., Senjem, M.L., Zeng, G., Bernstein, M.A., Gunter, J.L., Pankratz, V.S., Aisen, P.S., Weiner, M.W., Petersen, R.C., Shaw, L.M., Trojanowski, J.Q., Knopman, D.S., 2010. Brain beta-amyloid measures and magnetic resonance imaging atrophy both predict time-to-progression from mild cognitive impairment to Alzheimer's disease. Brain 133, 3336–3348. http://dx.doi. org/10.1093/brain/awq277. Kassubek, J., Müller, H.P., 2016. Computer-based magnetic resonance imaging as a tool in clinical diagnosis in neurodegenerative diseases. Expert Rev. Neurother. 16, 295–306. http://dx.doi.org/10.1586/14737175.2016.1146590. Kenny, E.R., Blamire, A.M., Firbank, M.J., O’Brien, J.T., 2012. Functional connectivity in cortical regions in dementia with Lewy bodies and Alzheimer’s disease. Brain 135, 569–581. http://dx.doi.org/10.1093/brain/awr327. Kinkingnéhun, S., Sarazin, M., Lehéricy, S., Guichart-Gomez, E., Hergueta, T., Dubois, B., 2008. VBM anticipates the rate of progression of Alzheimer disease: a 3-year longitudinal study. Neurology 70, 2201–2211. http://dx.doi.org/10.1212/01.wnl. 0000303960.01039.43. Kohonen, T., 1998. The self-organizing map. Neurocomputing 21, 1–6. http://dx.doi.org/ 10.1016/S0925-2312(98)00030-7. Lancaster, J.L., Woldorff, M.G., Parsons, L.M., Liotti, M., Freitas, C.S., Rainey, L.,

11

Journal of Neuroscience Methods xxx (xxxx) xxx–xxx

G. Singh et al.

earlier worked on devising standalone software for analysis of lipid molecules. His research interests include the design of new methods for image processing, image analysis, and machine learning. He is currently working on the development of Clinical Decision Support System (CDSS) for medical application.

http://dx.doi.org/10.1016/j.neuroimage.2007.09.073. Walker, L., McAleese, K.E., Thomas, A.J., Johnson, M., Martin-Ruiz, C., Parker, C., Colloby, S.J., Jellinger, K., Attems, J., 2015. Neuropathologically mixed Alzheimer’s and Lewy body disease: burden of pathological protein aggregates differs between clinical phenotypes. Acta Neuropathol 129, 729–748. http://dx.doi.org/10.1007/ s00401-015-1406-3. Xie, A., Gao, J., Xu, L., Meng, D., 2014. Shared mechanisms of neurodegeneration in Alzheimer’s disease and Parkinson’s disease. Biomed. Res. Int. 8 (2014). http://dx. doi.org/10.1155/2014/648740. Yang, W., Li, Y., Chen, X., 2014. ICA image feature extraction for improving diagnosis of Alzheimer’s disease and mild cognitive impairment. 2014 10th International Conference on Natural Computation (ICNC). IEEE 802–806. http://dx.doi.org/10. 1109/ICNC.2014.6975940. Zhang, D., Wang, Y., Zhou, L., Yuan, H., Shen, D., 2011. Multimodal classification of Alzheimer’s disease and mild cognitive impairment. Neuroimage 55, 856–867. http://dx.doi.org/10.1016/j.neuroimage.2011.01.008. Zhou, Y., Dougherty, J.H., Hubner, K.F., Bai, B., Cannon, R.L., Hutson, R.K., 2008. Abnormal connectivity in the posterior cingulate and hippocampus in early Alzheimer’s disease and mild cognitive impairment. Alzheimer’s Dement. 4, 265–270. http://dx.doi.org/10.1016/j.jalz.2008.04.006.

Lakshminarayanan Samavedham (Laksh) is currently an Associate Professor in the Department of Chemical and Biomolecular Engineering, National University of Singapore (NUS). His research interests have centered on machine learning and artificial intelligence with applications to chemical, biological and physiological processes/systems. He has authored/co-authored over 85 research articles in leading international scientific journals and has presented over 130 conference papers (including about 25 invited and keynote talks) in the field of process and medical systems. His current interests in teaching are in system dynamics, learning-centric use of technology and in the scale-up of education innovations. Erle Chuen-Hian Lim, MBBS, M Med (Int Med), FRCP (Glasgow) is Associate Professor, National University of Singapore and Senior Consultant Neurologist, National University Hospital. He completed his training in Neurology in 2000 and pursued a Movement Disorders fellowship with Prof C Warren Olanow and Mitchell F Brin at the Mount Sinai School of Medicine in New York, NY. Dr. Lim practices General Neurology with an interest in Movement Disorders. His interests include Parkinson’s disease and atypical parkinsonism, dystonia, Botulinum toxin (common and novel applications), and rating scales for movement disorders. His interests in general neurology include headaches and seizures.

Gurpreet Singh is presently working as cognitive software engineer for the Department of Radiology, Dalio Institute of Cardiovascular Imaging, NewYork-Presbyterian Hospital and the Weill Cornell Medicine, New York. He was a Ph.D. student in the Department of Chemical and Biomolecular Engineering, National University of Singapore (NUS). He has

12