1
A multi-stream convolutional neural network for
classification of progressive MCI in Alzheimer’s
disease using structural MRI images
Mona Ashtari-Majlan, Abbas Seifi, Mohammad Mahdi Dehshibi
Abstract — Early diagnosis of Alzheimer’s disease and
its prodromal stage, also known as mild cognitive impair-
ment (MCI), is critical since some patients with progressive
MCI will develop the disease. We propose a multi-stream
deep convolutional neural network fed with patch-based
imaging data to classify stable MCI and progressive MCI.
First, we compare MRI images of Alzheimer’s disease with
cognitively normal subjects to identify distinct anatomical
landmarks using a multivariate statistical test. These land-
marks are then used to extract patches that are fed into
the proposed multi-stream convolutional neural network to
classify MRI images. Next, we train the architecture in a
separate scenario using samples from Alzheimer’s disease
images, which are anatomically similar to the progressive
MCI ones and cognitively normal images to compensate for
the lack of progressive MCI training data. Finally, we trans-
fer the trained model weights to the proposed architecture
in order to fine-tune the model using progressive MCI and
stable MCI data. Experimental results on the ADNI-1 dataset
indicate that our method outperforms existing methods for
MCI classification, with an F1-score of 85.96%.
Index Terms — Alzheimer’s disease, Brain-shaped map,
Convolutional Neural Network, Multivariate statistical test,
Transfer learning.
I. INTRODUCTION
ALZHEIMER’S disease (AD) is a progressive neurode-
generative disorder that is one of the leading causes
of dementia in the elderly. According to [1], this disorder
affects over 30 million people worldwide. Early diagnosis
of this disease and its prodromal stage, also known as mild
cognitive impairment (MCI), is crucial since 10% to 15% of
MCI patients progress to AD, which is classified as progressive
MCI (pMCI) [2].
As AD advances, several brain regions develop structural
deformation and atrophy [3]. Structural Magnetic Resonance
Imaging (sMRI) is one of the most widely employed neu-
roimaging tools for predicting this disorder through identifying
brain atrophy [4] (see Fig. 1). In addition to sMRI (referred
to as MRI in this paper), non-invasive biomarkers such as (1)
M. Ashtri-Majlan is with the Department of Computer Sci-
ence, Universitat Oberta de Catalunya, Barcelona, Spain (e-mail:
ashtari.mona@gmail.com).
A. Seifi is with the Department of Industrial Engineering, Amirkabir
University of Technology, Tehran, Iran (e-mail: aseifi@aut.ac.ir).
M. M. Dehshibi is with the Department of Computer Science,
Universitat Oberta de Catalunya, Barcelona, Spain (e-mail: moham-
mad.dehshibi@yahoo.com).demographic information ( e.g., age and education) [5], and (2)
cognitive test scores [4] can also be used to provide possible
discriminative information for diagnosing AD in the early
stages. Several studies [4], [6], [7], [8], [9] have addressed
the MCI-to-AD conversion issue using neuroimaging methods
in conjunction with the biomarkers.
(a)
 (b)
Fig. 1. MRI samples of (a) Cognitively Normal (CN) and (b) AD
classes from Alzheimer’s Disease Neuroimaging Initiative database
(ADNI-1) [10]. In order to demonstrate the subtle brain atrophy, we
highlighted the affected regions.
Conventional methods in medical image processing typ-
ically use prior knowledge to segment brain images into
Regions of Interest (ROI) [11] or V oxel-Based Morphometry
(VBM) [12] to predict AD. While these methods can classify
stable MCI (sMCI) and pMCI, the focus was mainly on
Alzheimer’s disease. Deep learning algorithms, on the other
hand, have made it possible for researchers to combine feature
extraction, dimensionality reduction, and classification in an
end-to-end way. These algorithms also outperform conven-
tional methods in identifying AD patterns because they can
discover hidden representations among multiple regions of
neuroimages. Hence, these models have gained prominence
in analysing Alzheimer’s disease.
In this paper, we propose a multi-stream convolutional
neural network (CNN) for classifying sMCI and pMCI, which
is fed with patch-based imaging data extracted using a novel
data-driven technique. To accomplish so, we first use the
multivariate T2 Hotelling test to compare MRI images of AD
and cognitively normal (CN) individuals in order to identify
distinct anatomical landmarks. Following that, the statistical
test is performed on textural and statistical features extractedarXiv:2203.01944v1  [eess.IV]  3 Mar 20222
from MRI images. These landmarks are then used to generate
191919patches, with each landmark serving as the
centre of the patch. Finally, the extracted patches are fed
into the proposed multi-stream CNN to classify MRI images.
To compensate for the lack of pMCI training data, we first
train the proposed architecture using AD/CN images that are
anatomically similar to the pMCI ones. Then, we transfer the
weights of the trained model to the proposed architecture in
order to fine-tune the model using pMCI and sMCI data. The
contribution of this study is two-fold:
1) Rather than utilising non-rigid registration to identify
anatomical landmarks in the brain, we propose employ-
ing rigid registration to reduce computational complexity
and eliminate morphological structure deformations that
could cause inherent errors in the classification step.
Therefore, we partition each MRI image during the
anatomical landmark detection phase and perform the
T2 Hotelling test on these partitions to capture subtle
anatomical differences, to account for more information,
and to reduce the impact of potential errors caused by
inter-subject brain shape variations. Finally, the identi-
fied statistically significant landmarks centres from the
partitions are used as anchoring forces for selecting the
patches fed into the proposed multi-stream CNN.
2) We propose using transfer learning to classify
pMCI/sMCI classes in training the proposed architecture
to overcome the complexity of learning caused by the
subtle structural changes in MCI brains compared to AD
and CN. In the Ablation study, we also demonstrate the
importance of employing transfer learning. Furthermore,
we address intrinsic errors caused by inter-subject brain
shape differences by conducting experiments to deter-
mine an ideal image patch size in order to feed to the
proposed CNN model.
The rest of this paper is organised as follows: Section II
surveys the previous studies. Section III describes the proposed
method. Experimental results are given in Section IV. Finally,
the paper is concluded in Section V.
II. L ITERATURE REVIEW
Mild cognitive impairment (MCI) is a stage of cognitive
decline that occurs between the predicted cognitive loss as-
sociated with normal ageing and the more severe decline
associated with dementia. MCI may result in a higher level
of developing dementia caused by Alzheimer’s disease if the
anatomical changes in the brain are proactive. Progressive
MCI differs from stable MCI in the progression of functional
connectivity values over time. However, classifying pMCI and
sMCI patients is challenging due to the subtle anatomical
differences in the brain [13]. The four conventional feature
extraction approaches usually mentioned in the literature for
classifying pMCI and sMCI are voxel-based, slice-based, ROI-
based, and patch-based [14], although they are not entirely
mutually exclusive. In this section, before surveying recent
advances in deep learning-based methods for classifying pMCI
and sMCI, we will briefly review these four approaches by
discussing the advantages and disadvantages of each group.The voxel-based techniques [15], [16], [17] use the voxel
intensity values from all neuroimaging modalities. Although
voxel-based techniques are simple to implement, they typically
require spatial co-alignment of the input image to standard 3D
space and suffer from high dimension feature space compared
to available sample numbers. Ortiz et al. [18] used the t-
test algorithm to partition the brain area into 3D patches
in order to address the mentioned drawbacks and eliminate
non-significant voxels. The patches were then used to train
an ensemble of deep belief networks, and a voting scheme
was used to make the final prediction. However, as mentioned
by [19], there is an inherent over-fitting challenge with voxel-
based techniques.
The sliced-based techniques [20], [21] extract slices from
the 3D neuroimaging brain scan by projecting the sagittal,
coronal, and axial to the 2D image slices. Indeed, because
non-affected regions and normal slices must be chosen as the
reference distribution, they cannot account for the disease and
may be considered an anomaly [22]. Furthermore, choosing
separate 2D slices may neglect the spatial dependencies of
voxels in adjacent slices due to inter/intra anatomical variances
in the brain images [14]. However, sliced-based techniques
allow for the usage of a broader range of conventional and
deep learning-based approaches. For instance, different pre-
trained deep learning models on ImageNet, such as DenseNet,
VGG16, GoogLeNet, and ResNet, can be fine-tuned by 2D
slices to classify AD from CN [19]. In [21], researchers
extracted features from MRI image slices using a pre-trained
2D CNN and fed the extracted feature sequence to a recurrent
neural network (RNN). The RNN was in charge of determining
the relationship between the sequence of extracted features
corresponding to MRI image slices. However, sliced-based
techniques are computationally expensive due to the use of
additional learnable parameters which cannot directly benefit
from transfer learning.
ROI-based techniques consider brain regions that have
been predefined physically or functionally [23], [8], [24].
These methods use spatial information such as automated
anatomical labelling [25] and diffusion-weighted imaging in
MRI to extract features. The prominent regions that have
been considered by almost all ROI-based feature extraction
studies on AD prediction are the hippocampus, amygdala,
and entorhinal. However, one of the advantages of employing
ROI-based approaches is like a double-edged sword which can
be a disadvantage because ROI identification requires expert
human expertise. Furthermore, these techniques are considered
time-consuming due to the need for non-linear registration
and brain tissue segmentation. There is also the possibility
of information loss because the abnormal region may spread
from a single ROI to multiple ROIs.
Patched-based approaches [26], [27] partition the entire
brain into multiple patches from which numerous feature
vectors are extracted. The extracted patches include shapes,
texture, and volume features generated from distinct brain
regions or specific patterns. This computational approach
eliminates the need for manual ROI identification, and makes
it possible to use landmark-based patch extraction or other
discriminative patch-based biomarkers [8]. However, selecting3
useful patches from a complete image is problematic, mainly
due to increasing computational complexity when a non-
rigid registration approach is used. Researchers have used
rigid registration or embedded the registration step into deep
learning-based methods to address this challenge [28]. There
is another issue with patched-based approaches related to
multiple instance learning. Although this challenge is almost
addressed in classifying AD and CN by leveraging patch
relationships [29], [30], the difficulty in classifying pMCI
in Alzheimer’s disease is not still resolved. This issue is
linked to the bag’s labelling in classifying pMCI from sMCI,
where a bag can get a negative label even though it contains
informative anatomical landmark(s), but it cannot meet the
majority rule.
Convolutional neural networks (CNNs) are a popular deep
learning approach for classifying Alzheimer’s disease. Islam
and Zhang [31] proposed three 2D CNNs to generate three
distinct MRI views. Each CNN in their architecture comprised
three convolutional layers and four dense blocks, where the
final decision was made by majority vote. Oh et al. [32] pro-
posed a convolutional autoencoder-based approach for the AD
and CN classification task, addressing the pMCI data limitation
with transfer learning. Liu et al. [8] suggested a coarse-to-
fine hierarchical ensemble learning method for simultaneous
hippocampus segmentation and Alzheimer’s disease classi-
fication that employed a multi-task deep CNN architecture
and 3D densely connected CNNs. In this method, an MRI
image is first divided into multiple slices, and then a pre-
trained deep neural network is used to extract features from
the slices. The coarse predictions were then used in ensemble
learning to obtain refined results for all slices. Ebrahimi et
al. [21] extracted a sequence of features from 2D MRI slices
using a pre-trained ResNet18 [33], which they subsequently
trained a temporal convolutional network and several types
of RNNs to classify AD/CN. Zhao et al. [24] introduced aregion ensemble model with three sequential sub-networks
to account for a global feature map derived from the entire
brain and regional feature maps extracted using a segmentation
model. The feature representations were fused in their method,
and the classification was performed using an attention-based
technique. Researchers employed a data-driven technique to
select informative patches in [34], which resulted in specific
landmark localisation in brain MRI images. Each landmark
patch was then fed into the CNN models, which produced the
final classification result using the maximum voting strategy.
P¨olsterl et al. [35] proposed the dynamic affine feature map
transform, an auxiliary module for CNNs that dynamically
incites or represses each feature map of a convolutional layer
based on both image and tabular biomarkers. A more detailed
overview of deep learning algorithms for Alzheimer’s disease
classification can be found in [36], [14], [37].
III. P ROPOSED METHOD
Given a dataset of NsamplesD=f(xi; yi)gN
i=1, with xi2
Rdxandyi2Rdy, our goal is to train a multi-stream deep
convolutional neural network H(x) =E[YjX=x]to classify
sMCI and pMCI by minimising the cross-entropy between the
class labels and the softmax output as in Eq. 1
p(yijx;w; b) =exp(xTwi+bi)P
j2dyexp(xTwj+bj): (1)
where wandbare the network’s weights and bias terms, re-
spectively. In this study, we use the baseline 1.5T T1-weighted
MRI images of subjects from the ADNI-1 dataset [10], where
the input image Xhas a size of dx= 185155150and is
labelled by dy=f0;1g. The output label Yconsists of two
probability values in the [0;1]range withH(xi) = 0 if the i-th
sample belongs to the sMCI class and H(xi) = 1 otherwise.
It has been shown [38] that in the early stages of
Alzheimer’s disease, only certain brain regions are subject
(a)
Fig. 2. The schematic of the proposed multi-stream convolutional neural network.4
to morphological changes caused by the disease. Therefore,
we conduct a statistical test to identify these informative
landmarks in the MRI images and extract Lpatches from
each MRI image xi=fsi;jgL
j=1, with si;j2R191919.
The proposed data-driven approach for extracting patches from
the MRI image, on which a preprocessing step has been
performed, is described in the following subsections, followed
by the details of the multi-stream CNN. Figure 2 shows the
schematic of the proposed multi-stream architecture.
A. Preprocessing
We pre-process the MRI images to use them in the proposed
method. There are four steps in the preprocessing phase: (1)
anterior commissure-posterior commissure correction using
the 3D Slicer software1; (2) intensity inhomogeneity correction
using N4ITK [39], an enhanced version of nonparametric
nonuniform normalisation; (3) skull stripping using a pre-
trained U-Net2to remove both the skull and the dura; and
(4) rigid registration, which involves linearly aligning MRI
images to the Colin27 template and resampling them to a size
of155185150with a resolution of 111 mm3. Figure 3
shows a sample of MRI image from ADNI-1 dataset on which
the preprocessing is performed.
(a)
 (b)
 (c)
 (d)
Fig. 3. The visual representation of the preprocessing steps for an
MRI sample. (a) A raw MRI image, (b) the MRI image with anterior
commissure-posterior commissure correction (c) the MRI image with
intensity inhomogeneity correction, and (4) skull stripped MRI image.
The yellow line in (a) and (b) depicts the anterior commissure-posterior
commissure line.
B. Anatomical landmark detection
We must first identify the anatomical locations in the brain
that are most influenced by the disease before we can classify
sMCI and pMCI patients. As a result, we randomly divide
samples from AD and CN individuals into the train, validation,
and test sets with sizes of 0:7N,0:1N, and 0:2N,
respectively, where Nis the total number of samples. Then,
we select M= 0:7N(i.e., the training set) and propose a
novel data-driven landmark detection method in which MRI
images are partitioned into 555patches. As mentioned
in [40], [41], when a patient is diagnosed with Alzheimer’s
disease, several regions in the brain are subject to anatomical
degeneration. At the pMCI stage, the same regions undergo
anatomical changes, but the degeneration is not as severe as
those seen at the onset of Alzheimer’s disease. With respect to
this fact, we use identical anatomical locations for classifying
sMCI and pMCI patients.
1http://www.slicer.org/
2https://github.com/iitzco/deepbrainEach partition is then represented by a 29-dimensional
feature vector. This feature vector includes the Gray-Level Co-
Occurrence Matrix (GLCM) [42], Structural Similarity Index
Measure (SSIM) [43], Mean Square Error (MSE), entropy, and
the mean and standard deviation of the partition voxels. To
extract the GLCM elements of the feature vector, we generate
six GLCM matrices with three adjacency directions, namely
horizontal, vertical, and in-depth, each of which is associated
with two distance values. Then, we extract contrast, corre-
lation, homogeneity, and entropy, from each GLCM matrix
resulting in a total of 24 elements. The Colin27 template [44]
is used as the reference image for measuring SSIM and MSE
at each patch location. Finally, we apply the multivariate T2
Hotelling statistical test [45] to generate a brain-shaped p-
value map (see Fig. 4). Algorithm 1 details the steps of
generating the brain-shaped p-value map.
Fig. 4. A brain-shaped p-value map in which top 50 landmark locations
are represented by spheres in a gradient colour from red to blue, with p-
values ranging from 0 to 0.001. Each p-value is paired with a landmark
(lx; ly; lz).
After obtainingPset, we exclude landmarks with a spatial
Euclidean distance of less than 15 to reduce the redundancy of
overlapped adjacent patches to identify the most discriminative
anatomical landmarks in the brain. The top 50 landmarks with
the lowest p-values are then chosen (see Fig. 4).
The registration and landmark detection steps in Algo-
rithm 1 are affected by the MRI image partitioning size, i.e.,
555. In the case of selecting a smaller partition size,
the lack of adequate morphological variations could lead to
discarding informative landmarks. In contrast, larger partition
sizes cause intrinsic physiological differences to eclipse subtle
disease-related changes.
Therefore, after obtaining the top 50 landmarks, we sample
27 3D image patches with a 333displacement around
the centre of each landmark to increase the size of patches
to191919with two intentions: (i) compensating for
regions that may unintentionally be discarded in anatomical
landmark detection, and (ii) providing sufficient morphological
structures for each stream of the proposed CNN to construct
a discriminative latent space.5
Algorithm 1: Generating brain-shaped p-value map.
Input : AD=fxAD
1; xAD
2;


; xAD
Mg
CN=fxCN
1; xCN
2;


; xCN
Mg
K= 34;410, which is the total number of
patches with a size of 555.
Output:P: A set of p-value , forming the brain-shaped
p-value map.
Step 1: Partitioning
1VAD Partition( AD);
//VAD=fp1;1;


; p1;K;


; pM;1;


; pM;Kg.
2VCN Partition( CN);
//VCN=fq1;1;


; q1;K;


; qM;1;


; qM;Kg.
Step 2: Feature extraction & T2 Hotelling test
3forj 1toKdo
4 fori 1toMdo
5 fAD
i;j Feature-Extraction( pi;j);
6 fCN
i;j Feature-Extraction( qi;j);
7 end
8 pvalue j Hotelling-Test( fAD
M;j; fCN
M;j);
//fM;jis a M29matrix,
representing extracted features
form the jthpatch.
9end
10P Sort( pvalue; asc);
// Each pvalue is paired with a
landmark (lx; ly; lz).
11returnP
C. Multi-stream classifier architecture
We propose a multi-stream CNN architecture with L
streams, each fed with the patch si;jextracted from the input
image xiand centred on the identified landmark location
(lx;j; ly;j; lz;j). We construct the patch si;jwith a size of
191919, surrounding the corresponding landmark location,
in order to better represent morphological variations in the
MRI images. The local spectral-spatial feature is extracted
from each 3D image patch by each stream of the proposed
CNN architecture.
As depicted in Fig. 2, the proposed multi-stream CNN has
L= 50 streams, with an identical structure. Each stream has
five convolutional layers (Conv), followed by a rectified linear
unit (ReLU) activation function. The convolutional layers have
32, 64, 64, 128, and 128 333convolution filters,
respectively. After Conv2, Conv4, and Conv5, we consider
batch normalisation and 222max-pooling layers. There
are three fully connected layers with 128, 64, and 8 units at the
end of each stream, which are followed by a dropout layer with
a ratio of 0.4 to prevent overfitting. Although the architecture
of all 50 streams is the same, their weights are tuned and
updated separately, where the input patches for each stream
are randomly selected to avoid the unintentional bias towards
ordering streams. We concatenate the outputs of 50 streams
and add a dropout layer with a ratio of 0.6 to fuse the locally
extracted spectral-spatial features. Before passing the featurevector into the softmax function for the final classification,
we add three fully connected layers with 64, 64, and 32 units,
respectively.
IV. E XPERIMENTS
A. ADNI-1 dataset
In this study, we use the baseline 1.5T T1-weighted MRI
images of subjects from the ADNI-1 dataset [10]. The vol-
umetric 3D MPRAGE protocol is used to acquire sagittal
T1-weighted MRI images with an in-plane spatial resolution
of1:251:25 mm2and 1.2 mm thick sagittal slices. The
imaging dataset contains baseline images from 695 partic-
ipants including 200 Alzheimer’s disease, 231 cognitively
normal, 164 progressive MCI, and 100 stable MCI. Figure 5
shows four samples from this dataset, and Table I presents the
demographic and clinical information of subjects in ADNI-1.
(a)
 (b)
 (c)
 (d)
Fig. 5. Four samples from ADNI-1 dataset [10] (a) AD, (b) CN, (c) pMCI,
and (d) sMCI
B. Architecture details and Evaluation metrics
The proposed architecture is implemented using Python
based on the Keras package3, on a computer with Intel(R)
Core(TM) i7-4790K @4.00 GHz CPU and 16G RAM. We
trained the network using Adam optimiser [46] with the first
momentum of 0.9 and the second momentum of 0.999. The
initial learning rate and the constant for numerical stability
are set to 103and106, respectively. We set the maximum
number of training epochs to 40 and used a mini-batch-size of
5 at each iteration, where the training data was shuffled before
each training epoch. There are two other hyper-parameters
which are the number of streams Land the patch size.
We evaluate the performance of the proposed ar-
chitecture with the number of streams in the range
f10;20;30;40;50;60gand the size of patches in the range
f9;11;15;19;23gusing the validation set from the AD and
CN classes.It is worth noting that because there are two
hyperparameters, one must be fixed in order to generate the
other. Figure 6 shows the accuracy of the proposed multi-
stream CNN as a function of these two hyper-parameters.
While increasing the number of streams improves classifi-
cation accuracy (see Fig. 6a), it also increases computational
complexity. We can also observe in Fig. 6b that a larger patch
size ( i.e.,232323) shows a performance drop when
compared with the 191919patch size. This is owing
to the fact that enlarging the patch size includes tissues in that
3https://github.com/fchollet/keras6
TABLE I
DEMOGRAPHIC AND CLINICAL INFORMATION OF SUBJECTS IN ADNI-1. V ALUES ARE REPORTED AS MEAN STANDARD DEVIATION .
Class Male/Female Age ADAS CDR-sb FAQ MMSE NPI-Q
AD 103/97 75.64 7.71 13.02 5.23 4.39 1.6 13.16 6.71 23.31 2.03 3.44 3.27
CN 119/112 76.155 4.99 28.51 4.89 0.03 0.12 0.13 0.59 29.09 0.98 0.36 0.95
sMCI 66/34 75.44 7.27 22.93 5.78 1.24 0.62 1.65 3.00 27.65 1.70 1.45 2.40
pMCI 97/67 74.54 7.05 17.67 5.14 1.87 0.96 5.64 5.15 26.62 1.71 2.30 3.11
ADAS: Alzheimer’s Disease Assessment Scale
CDR-sb: Clinical Dementia Rating ‘sum of boxes’
FAQ: Functional Activities QuestionnaireMMSE: Mini-Mental State Examination
NPI-Q: Neuropsychiatric Inventory Questionnaire
(a)
 (b)
Fig. 6. The accuracy of the proposed multi-stream CNN classifier as a
function of (a) the number of streams, and (b) the patch size.
patch of the brain whose intrinsic morphological similarity
is dominant to subtle changes that aid in the early diagnosis
of the disease. Therefore, instead of discriminating the subtle
changes induced by the disease’s development in the early
stages, the network discriminates irrelevant tissues.As a result,
to establish a trade-off between computational complexity and
accuracy, we set the number of streams Lto 50 and the patch
size to 191919and preserve these values in the rest of
the experiments.
To evaluate the performance of the proposed architecture,
we use the metrics as in Eq. 2.
ACC =TP+TN
TP+TN+FP+FN:
SEN =TP
TP+FN;
SPE =TN
TN+FP;
F1score =2SENSPE
SEN + SPE:(2)
where TP,TN,FP, and FN denote true positive, true
negative, false positive, and false negative, respectively. We
also report area under receiver operating characteristic curve
(AUC).
C. Experimental results
To assess the performance of the proposed architecture,
we performed three steps: (1) training the proposed multi-
stream CNN with MRI images of AD and CN subjects, (2)
transferring weights from the trained model in Step 1 to the
identical architecture and fine-tuning it with the data related
to sMCI and pMCI patients, and (3) adding biomarkers asan auxiliary modality to examine the reciprocal influence of
spectral-spatial features and biomarkers.
(1) Classification of AD and CN subjects : To train the pro-
posed multi-stream CNN, we randomly select 70% of the MRI
patches from the two classes of AD and CN as the training
set, 20% as the test set, and 10% as the validation set. We
also compare the trained model in this step with 10 additional
approaches, including regions of interest (ROI) [47], V oxel-
Based Morphometry (VBM) [12], and five deep learning-based
methods. We used the FAST algorithm [48] to segment brain
MRI images into three different tissues for the ROI and VBM
comparison: White Matter (WM), Gray Matter (GM), and
Cerebrospinal Fluid (CSF). We followed the implementations
described by the researchers for the deep learning-based
approaches.
For the RIO-based method, we align the anatomical auto-
mated labelling template [25] with the native space of each
MRI image. This template contains 116 predefined regions
in which the extracted GM tissue volumes are normalised
using the summation of GM, WM, and CSF volumes. This
normalised tissue volume is used as the representative feature
for the MRI images. For the VBM method, we use affine reg-
istration with 12 degrees of freedom to align MRI images with
the Colin27 template [44] in order to extract the GM density
as the representative feature. To reduce the dimensionality of
this feature vector, we perform the t-student statistical test on
the extracted features from AD and CN subjects and chose
GM densities with p-values less than 0.001.
Finally, we classify AD and CN using these two feature
representations using a linear support vector machine
(SVM) [50] with soft-margin and a multilayer perceptron
(MLP). The MLP comprises two hidden layers, with 13
and 15 neurons, and a binary output layer with a logistic
activation function. We train these classifiers using the 5-fold
cross-validation strategy. To ensure a fair comparison for
MCI conversion prediction, we train the VBM and ROI
models on AD/CN and test them on an independent test set
of pMCI/sMCI.
(2) Classifying sMCI and pMCI using transfer learning :
The availability of MRI images for sMCI and pMCI patients
is less than that of AD because the symptoms of Alzheimer’s
disease are not as severe at these stages. As a result, many
patients may not be asked to get an MRI test. Furthermore,
structural changes in MCI brains caused by dementia may be7
TABLE II
RESULTS OF CLASSIFYING AD/CN AND P MCI/ SMCI USING THE PROPOSED MULTI -STREAM CNN S ALONG WITH 10ADDITIONAL APPROACHES .
MethodAD vs. CN (%) pMCI vs. sMCI (%)
ACC SEN SPE F1-score AUC ACC SEN SPE F1-score AUC
ROI + SVM 70.30 68.51 67.50 67.83 79.39 59.47 66.86 68.90 67.87 60.99
VBM + SVM 82.84 82.54 80.50 81.34 90.39 70.08 81.48 67.07 73.58 74.38
ROI + MLP 73.08 70.00 71.19 70.59 77.52 58.75 70.00 66.04 67.96 53.93
VBM + MLP 83.85 78.33 85.45 81.74 89.92 70.00 72.00 78.26 75.00 73.93
Shmulev et al. [49] - - - - - 76.00 70.00 88.00 77.97 86.00
DM2L[5] 91.09 93.50 88.05 90.69 95.86 76.90 82.43 42.11 55.74 77.64
H-FCN [6] 90.30 96.50 82.40 85.08 95.10 80.90 85.40 52.60 65.10 78.10
HybNet [28] 91.90 82.40 94.50 91.50 96.50 82.70 57.90 86.60 69.39 79.30
Zhao et al. [24] - - - - - 85.90 50.00 91.60 64.68 85.40
Proposed architectureLbiomarkers 97.54 95.54 99.40 97.43 99.38 69.21 68.56 93.15 78.98 77.35
Proposed architecture 97.78 95.59 99.82 97.66 99.97 79.90 75.55 99.70 85.96 94.39
very subtle compared to CN and AD, making the convergence
of the proposed multi-stream CNN challenging. To overcome
these limitations, we first trained our model using data from
CN and AD MRI images and fine-tune the model with data
gathered from sMCI and pMCI patients.
Since the pre-trained streams of the proposed architecture
are designed to extract structural features that correspond to
AD, we freeze the initial layers and only re-train the last
three fully connected layers. To fine-tune the model, we used
70% of the pMCI and sMCI MRI patches and tested the
fine-tuned model with the remaining data.
(3) Unleashing the impact of biomarkers : In addition to
MRI images, non-invasive biomarkers such as demographic
information and cognitive test scores [4] can also be used to
provide potentially discriminatory information for diagnosing
Alzheimer’s disease in its early stages. We propose adding
biomarkers as an auxiliary modality to the architecture in order
to evaluate the reciprocal influence of spectral-spatial features
and biomarkers. Therefore, we introduce an additional input
stream to the proposed architecture to incorporate numerical
values offAge, ADAS, CDR-sb, FAQ, MMSE, NPI-Q g(see
Table I). However, a subset of the MRI images in the ADNI-
1 dataset does not include biomarkers, resulting in missing
values when incorporating them into the proposed CNN archi-
tecture. We use k-Nearest Neighbour (kNN) as a preprocessing
step to handle missing values. First, we execute the kNN
with k= 6 (analogous to 6 biomarkers) on the available
data. Then, for those entries that do not have a biomarker,
we fill in the value with the average values of the knearest
neighbours. Finally, we incorporate these biomarkers into the
architecture at the ‘Concatenate’ layer while preserving the
learning parameters as described in Section IV-B.
Table II shows the performance of the proposed architec-
tures compared with 10 other approaches. As shown in Ta-
ble II, the proposed multi-stream CNNs achieve considerably
better F1-scores than both conventional and deep learning-
based approaches. While the biomarkers can improve the
classification performance, the lack of large-scale data is
a barrier to incorporating them into the proposed method.
Furthermore, as previously stated, the cognitive test scores
of individuals suffering from various types of MCI differ
slightly. This subtle difference could explain why the multi-stream architecture combined with biomarkers performs poorly
in distinguishing pMCI and sMCI when compared with the AD
and CN classification. We also plot the ROC curves in Fig. 7
for classifying both AD vs. CN and pMCI vs. sMCI.
(a)
 (b)
Fig. 7. ROC curves of the proposed multi-stream architecture, the
proposed architecture integrated with biomarkers, ROI, and VBM-based
methods for (a) AD/CN classification, and (b) pMCI/sMCI classification.
In the case of classifying AD and CN, Fig. 7a shows that
the proposed architecture achieves an AUC of 99.97%, outper-
forming ROI, VBM, and multi-stream architecture integrated
with biomarkers, which achieve AUCs of 79.39%, 90.39%,
and 99.38%, respectively. Furthermore, as demonstrated in
Fig. 7b, when we use transfer learning, the fine-tuned model
can provide appropriate discriminative information to classify
pMCI and sMCI subjects. This evidence can also convey that
AD and pMCI have structural deformations similar to the
extracted features from various landmarks in different streams
of the proposed architecture.
The results in Table II show that the proposed multi-stream
architecture combined with biomarkers is slightly less efficient
than the multi-stream CNN in distinguishing pMCI from
sMCI when compared with the AD and CN classification. To
highlight these subtle differences and better comprehend the
reciprocal influence of spectral-spatial features and biomark-
ers, we have visualised the class-discriminative localisation
map and the latent space in Fig. 8 using Grad-CAM [51] and
t-SNE [52], respectively. Figure 8a shows an example of a
class-discriminative localisation map for the last convolutional
layers in the proposed multi-stream CNN in which the salient
regions of each brain patch have been identified so that the
classifier could distinguish between AD/CN and pMCI/sMCI8
(a)
 (b)
Fig. 8. (a) An example of a class-discriminative localisation map for the last convolutional layers in the proposed multi-stream CNN using Grad-
CAM [51]. Each stream has 128 filters. However, in order to keep the publication page limits, we only illustrate five out of 50 patches, each
representing five filters. The colourbar represents the relevance of each pixel in the Grad-CAM, where the red colour has the highest importance
in the trained network’s final decision and the blue colour has the least importance. (b) Visualisation of the feature space using t-SNE [52] for the
classification of AD/CN and pMCI/sMCI with and without biomarkers.
with high potential. Furthermore, we have observed in Fig. 8b
that introducing biomarkers as an auxiliary modality causes the
clusters ( i.e., AD/CN and pMCI/sMCI) to become tighter with
a degree of overlap than without the biomarkers. This evidence
can explain why the proposed multi-stream CNN without
biomarkers performs marginally better, where the latent space
provides for clean cluster separation.
D. Ablation study
In the ablation study, we conduct two sets of experiments
to better understand the efficacy of the proposed multi-stream
CNN for the classification of pMCI in Alzheimer’s disease.
We therefore examine:
— Contribution of transfer learning : In this experiment,
instead of using transfer learning, we test the baseline perfor-
mance when the model is directly trained on pMCI/sMCI data.
We have 100 and 164 sMCI and pMCI samples, respectively,
and use the same data split as in the transfer learning (70/30)
for a fair comparison. The other learning parameters are
preserved as described in Section IV-B. The evaluation metrics
(i.e., ACC: 63.53%, SEN: 100%, SPE: 0.0%, and F1-score:
77.70%) reveal that the model cannot converge to classify
pMCI/sMCI classes without using transfer learning. In fact,
the trained model predicts the same optimal output regardless
of the input, with the average answer minimising loss.
— Contribution of multi-stream architecture : In this ex-
periment, instead of the proposed multi-stream CNN, we use
a single stream CNN to which 3D patches are fed. The
model is trained on the AD/CN dataset and fine-tuned using
pMCI/sMCI data with the same data split used in the previous
experiments. The evaluation metrics ( i.e., ACC: 56.96%, SEN:
67.35%, SPE: 40.00%, and F1-score: 66.00%) show that the
model is unable to converge to classify pMCI/sMCI classes,
and the trained model is predicting random class for all the
data points regardless of the input. These metrics highlighttwo main facts: (1) with the multi-stream architecture, the
model can build a more discriminative latent space as a result
of focusing on the landmarks identified in the anatomical
landmark detection phase, which are more likely to be signs
of Alzheimer’s disease. However, with a single-stream ar-
chitecture, the model cannot distinguish between anatomical
abnormalities caused by the disease and those caused by the
morphology of the brain. Indeed, the model’s sensitivity and
specificity are inadvertently changed due to the morphological
similarity of the inputs; (2) input patches to the single-stream
architecture predominantly capture local information of the
image, and the global relationship between different landmark
locations is no longer taken into account.
V. C ONCLUSION
Alzheimer’s disease is the most common type of dementia,
resulting in memory impairment and cognitive decline. Mild
cognitive impairment is a prodromal stage of AD, also known
as the transition stage. MCI patients either progress to AD or
remain at the same stage over time. Therefore, it is critical to
distinguish between progressive MCI and stable MCI in early
stages to prevent rapid progression of the disease.
In this study, we have proposed a method for classifying
pMCI and sMCI patients using MRI images. The proposed
method consists of two main steps. (1) We have developed
a novel data-driven approach based on the multivariate T2
Hotelling statistical test to identify anatomical landmarks in
MRI images and generate a brain-shaped p-value map. Each
landmark is paired with a 3D coordinate, allowing us to extract
patches of 191919. (2) We have proposed a multi-stream
deep convolutional neural network in which each stream is
fed by one of the patches. This multi-stream CNN employed
transfer learning to classify pMCI and sMCI patients using the
ADNI-1 dataset. We assessed the proposed method in three
experimental steps and demonstrated the significance of our
contributions to transfer learning and multi-stream architecture9
in the Ablation study. We have performed several experiments
to evaluate the efficiency of the proposed architecture based
on the best practices. Experimental results have shown that
our method outperformed existing approaches, particularly in
the classification of MCI patients. Thus, this method can
assist practitioners to expand investigating on various diseases
associated with structural atrophy.
REFERENCES
[1] D. E. Barnes and K. Yaffe, “The projected effect of risk factor reduction
on alzheimer’s disease prevalence,” The Lancet Neurology , vol. 10, no. 9,
pp. 819–828, 2011.
[2] R. C. Petersen, R. O. Roberts, D. S. Knopman, B. F. Boeve, Y . E.
Geda, R. J. Ivnik, G. E. Smith, and J. Jack, Clifford R., “Mild Cognitive
Impairment: Ten Years Later,” Archives of neurology , vol. 66, no. 12,
pp. 1447–1455, 2009.
[3] M. A. DeTure and D. W. Dickson, “The neuropathological diagnosis of
Alzheimer’s disease,” Molecular neurodegeneration , vol. 14, no. 1, pp.
1–18, 2019.
[4] K. Hett, V .-T. Ta, I. Oguz, J. V . Manj ´on, and P. Coup ´e, “Multi-scale
graph-based grading for alzheimer’s disease prediction,” Medical Image
Analysis , vol. 67, pp. 101 850: 1–13, 2021.
[5] M. Liu, J. Zhang, E. Adeli, and D. Shen, “Joint classification and
regression via deep multi-task multi-channel learning for alzheimer’s
disease diagnosis,” IEEE Transactions on Biomedical Engineering ,
vol. 66, no. 5, pp. 1195–1206, 2019.
[6] C. Lian, M. Liu, J. Zhang, and D. Shen, “Hierarchical Fully Convolu-
tional Network for Joint Atrophy Localization and Alzheimer’s Disease
Diagnosis Using Structural MRI,” IEEE transactions on pattern analysis
and machine intelligence , vol. 42, no. 4, pp. 880–893, 2020.
[7] Q. Li, X. Xing, Y . Sun, B. Xiao, H. Wei, Q. Huo, M. Zhang, X. S. Zhou,
Y . Zhan, Z. Xue, and F. Shi, “Novel Iterative Attention Focusing Strategy
for Joint Pathology Localization and Prediction of MCI Progression,”
inMedical Image Computing and Computer Assisted Intervention –
MICCAI 2019 . Springer International Publishing, 2019, pp. 307–315.
[8] M. Liu, F. Li, H. Yan, K. Wang, Y . Ma, L. Shen, and M. Xu, “A multi-
model deep convolutional neural network for automatic hippocampus
segmentation and classification in Alzheimer’s disease,” NeuroImage ,
vol. 208, pp. 116 459: 1–15, 2020.
[9] S. Qiu, P. S. Joshi, M. I. Miller, C. Xue, X. Zhou, C. Karjadi, G. H.
Chang, A. S. Joshi, B. Dwyer, S. Zhu, M. Kaku, Y . Zhou, Y . J. Alderazi,
A. Swaminathan, S. Kedar, M.-H. Saint-Hilaire, S. H. Auerbach, J. Yuan,
E. A. Sartor, R. Au, and V . B. Kolachalama, “Development and
validation of an interpretable deep learning framework for alzheimer’s
disease classification,” Brain , vol. 143, no. 6, pp. 1920–1933, 2020.
[10] C. R. Jack Jr., M. A. Bernstein, N. C. Fox, P. Thompson, G. Alexander,
D. Harvey, B. Borowski, P. J. Britson, J. L. Whitwell, C. Ward, A. M.
Dale, J. P. Felmlee, J. L. Gunter, D. L. Hill, R. Killiany, N. Schuff,
S. Fox-Bosetti, C. Lin, C. Studholme, C. S. DeCarli, G. Krueger, H. A.
Ward, G. J. Metzger, K. T. Scott, R. Mallozzi, D. Blezek, J. Levy, J. P.
Debbins, A. S. Fleisher, M. Albert, R. Green, G. Bartzokis, G. Glover,
J. Mugler, and M. W. Weiner, “The alzheimer’s disease neuroimaging
initiative (adni): Mri methods,” Journal of Magnetic Resonance Imaging ,
vol. 27, no. 4, pp. 685–691, 2008.
[11] C. Cabral, P. M. Morgado, D. Campos Costa, and M. Silveira, “Predict-
ing conversion from mci to ad with fdg-pet brain images at different
prodromal stages,” Computers in Biology and Medicine , vol. 58, pp.
101–109, 2015.
[12] J. Ashburner and K. J. Friston, “V oxel-based morphometry—the meth-
ods,” Neuroimage , vol. 11, no. 6, pp. 805–821, 2000.
[13] U. R. Acharya, S. L. Fernandes, J. E. WeiKoh, E. J. Ciaccio, M. K. M.
Fabell, U. J. Tanik, V . Rajinikanth, and C. H. Yeong, “Automated De-
tection of Alzheimer’s Disease Using Brain MRI Images– A Study with
Various Feature Extraction Techniques,” Journal of Medical Systems ,
vol. 43, no. 9, pp. 1–14, 2019.
[14] X. Zhao, C. K. E. Ang, U. R. Acharya, and K. H. Cheong, “Application
of Artificial Intelligence techniques for the detection of Alzheimer’s
disease using structural MRI images,” Biocybernetics and Biomedical
Engineering , vol. 41, no. 2, pp. 456–473, 2021.
[15] S. Basheera and M. S. S. Ram, “A novel CNN based Alzheimer’s
disease classification using hybrid enhanced ICA segmented gray matter
of MRI,” Computerized Medical Imaging and Graphics , vol. 81, pp.
101 713: 1–15, 2020.[16] J. Ruiz, M. Mahmud, M. Modasshir, and M. S. Kaiser, “3d densenet
ensemble in 4-way classification of alzheimer’s disease,” in 13th In-
ternational Conference on Brain Informatics . Springer International
Publishing, 2020, pp. 85–96.
[17] W. Li, X. Lin, and X. Chen, “Detecting alzheimer’s disease based on 4d
fmri: An exploration under deep learning framework,” Neurocomputing ,
vol. 388, pp. 280–287, 2020.
[18] A. Ortiz, J. Munilla, J. M. G ´orriz, and J. Ram ´ırez, “Ensembles of deep
learning architectures for the early diagnosis of the alzheimer’s disease,”
International journal of neural systems , vol. 26, no. 07, p. 1650025,
2016.
[19] M. A. Ebrahimighahnavieh, S. Luo, and R. Chiong, “Deep learning to
detect Alzheimer’s disease from neuroimaging: A systematic literature
review,” Computer Methods and Programs in Biomedicine , vol. 187, p.
105242, 2020.
[20] Y . Zhang, Q. Teng, Y . Liu, Y . Liu, and X. He, “Diagnosis of alzheimer’s
disease based on regional attention with smri gray matter slices,” Journal
of Neuroscience Methods , vol. 365, pp. 109 376: 1–8, 2022.
[21] C. R. A. D. N. I. Ebrahimi A, Luo S, “Deep sequence modelling for
alzheimer’s disease detection using mri,” Computers in Biology and
Medicine , p. 104537, 2021.
[22] R. Dey and Y . Hong, “Asc-net: Adversarial-based selective network
for unsupervised anomaly segmentation,” in International Conference
on Medical Image Computing and Computer-Assisted Intervention .
Springer, 2021, pp. 480–489.
[23] V . S. Rallabandi, K. Tulpule, M. Gattu, A. D. N. Initiative et al. , “Au-
tomatic classification of cognitively normal, mild cognitive impairment
and Alzheimer’s disease using structural MRI analysis,” Informatics in
Medicine Unlocked , vol. 18, pp. 100 305: 1–7, 2020.
[24] Y .-X. Zhao, Y .-M. Zhang, M. Song, and C.-L. Liu, “Region Ensemble
Network for MCI Conversion Prediction with a Relation Regularized
Loss,” in International Conference on Medical Image Computing and
Computer-Assisted Intervention . Springer, 2021, pp. 185–194.
[25] N. Tzourio-Mazoyer, B. Landeau, D. Papathanassiou, F. Crivello,
O. Etard, N. Delcroix, B. Mazoyer, and M. Joliot, “Automated Anatom-
ical Labeling of Activations in SPM Using a Macroscopic Anatomi-
cal Parcellation of the MNI MRI Single-Subject Brain,” Neuroimage ,
vol. 15, no. 1, pp. 273–289, 2002.
[26] M. Liu, J. Zhang, D. Nie, P.-T. Yap, and D. Shen, “Anatomical
landmark based deep feature representation for mr images in brain
disease diagnosis,” IEEE journal of biomedical and health informatics ,
vol. 22, no. 5, pp. 1476–1485, 2018.
[27] J. Wen, E. Thibeau-Sutre, M. Diaz-Melo, J. Samper-Gonz ´alez,
A. Routier, S. Bottani, D. Dormont, S. Durrleman, N. Burgos,
and O. Colliot, “Convolutional neural networks for classification of
alzheimer’s disease: Overview and reproducible evaluation,” Medical
Image Analysis , vol. 63, pp. 101 694: 1–19, 2020.
[28] C. Lian, M. Liu, Y . Pan, and D. Shen, “Attention-guided hybrid network
for dementia diagnosis with structural mr images,” IEEE Transactions
on Cybernetics , 2020.
[29] M. Liu, J. Zhang, E. Adeli, and D. Shen, “Landmark-based deep multi-
instance learning for brain disease diagnosis,” Medical image analysis ,
vol. 43, pp. 157–168, 2018.
[30] M. Ilse, J. Tomczak, and M. Welling, “Attention-based Deep Multi-
ple Instance Learning,” in 35th International Conference on Machine
Learning . PMLR, 2018, pp. 2127–2136.
[31] J. Islam and Y . Zhang, “Brain MRI analysis for Alzheimer’s disease
diagnosis using an ensemble system of deep convolutional neural
networks,” Brain Informatics , vol. 5, no. 2, pp. 1–14, 2018.
[32] K. Oh, Y .-C. Chung, K. W. Kim, W.-S. Kim, and I.-S. Oh, “Classification
and Visualization of Alzheimer’s Disease using V olumetric Convolu-
tional Neural Network and Transfer Learning,” Scientific Reports , vol. 9,
no. 1, pp. 1–16, 2019.
[33] K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for
Image Recognition,” in 2016 IEEE Conference on Computer Vision and
Pattern Recognition (CVPR) , 2016, pp. 770–778.
[34] M. Liu, C. Lian, and D. Shen, “Anatomical-landmark-based deep learn-
ing for alzheimer’s disease diagnosis with structural magnetic resonance
imaging,” in Deep Learning in Healthcare: Paradigms and Applications .
Springer International Publishing, 2020, pp. 127–147.
[35] S. P ¨olsterl, T. N. Wolf, and C. Wachinger, “Combining 3d image
and tabular data via the dynamic affine feature map transform,” in
International Conference on Medical Image Computing and Computer-
Assisted Intervention . Springer, 2021, pp. 688–698.
[36] N. Goenka and S. Tiwari, “Deep learning for alzheimer prediction using
brain biomarkers,” Artificial Intelligence Review , vol. 54, no. 7, pp. 1–45,
2021.10
[37] G. Chamarajan, Y . Charishma et al. , “Alzheimer’s disease: A survey,”
International Journal of Artificial Intelligence , vol. 8, no. 1, pp. 33–39,
2021.
[38] H. Braak, E. Braak, D. Yilmazer, R. A. I. De V os, E. N. H. Jansen,
and J. Bohl, “Pattern of brain destruction in parkinson’s and alzheimer’s
diseases,” Journal of Neural Transmission , vol. 103, no. 4, pp. 455–490,
1996.
[39] N. J. Tustison, B. B. Avants, P. A. Cook, Y . Zheng, A. Egan, P. A.
Yushkevich, and J. C. Gee, “N4itk: Improved n3 bias correction,” IEEE
Transactions on Medical Imaging , vol. 29, no. 6, pp. 1310–1320, 2010.
[40] Y . Fan, N. Batmanghelich, C. M. Clark, and C. Davatzikos, “Spatial
patterns of brain atrophy in MCI patients, identified via high-dimensional
pattern classification, predict subsequent cognitive decline,” NeuroIm-
age, vol. 39, no. 4, pp. 1731–1743, 2008.
[41] M. L. Schroeter, T. Stein, N. Maslowski, and J. Neumann, “Neural corre-
lates of alzheimer’s disease and mild cognitive impairment: A systematic
and quantitative meta-analysis involving 1351 patients,” NeuroImage ,
vol. 47, no. 4, pp. 1196–1206, 2009.
[42] R. M. Haralick, K. Shanmugam, and I. Dinstein, “Textural features
for image classification,” IEEE Transactions on Systems, Man, and
Cybernetics , vol. SMC-3, no. 6, pp. 610–621, 1973.
[43] Z. Wang, A. Bovik, H. Sheikh, and E. Simoncelli, “Image quality assess-
ment: from error visibility to structural similarity,” IEEE Transactions
on Image Processing , vol. 13, no. 4, pp. 600–612, 2004.
[44] C. J. Holmes, R. Hoge, L. Collins, R. Woods, A. W. Toga, and
A. C. Evans, “Enhancement of mr images using registration for signal
averaging,” Journal of computer assisted tomography , vol. 22, no. 2, pp.
324–333, 1998.
[45] H. Hotelling, The Generalization of Student’s Ratio . Springer New
York, 1992, pp. 54–65.
[46] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
inThird International Conference on Learning Representations, ICLR ,
2015, pp. 1—-15.
[47] D. Zhang, Y . Wang, L. Zhou, H. Yuan, D. Shen, A. D. N. Initiative et al. ,
“Multimodal classification of alzheimer’s disease and mild cognitive
impairment,” NeuroImage , vol. 55, no. 3, pp. 856–867, 2011.
[48] S. M. Smith, M. Jenkinson, M. W. Woolrich, C. F. Beckmann, T. E.
Behrens, H. Johansen-Berg, P. R. Bannister, M. De Luca, I. Drobnjak,
D. E. Flitney, R. K. Niazy, J. Saunders, J. Vickers, Y . Zhang, N. De Ste-
fano, J. M. Brady, and P. M. Matthews, “Advances in functional and
structural mr image analysis and implementation as fsl,” NeuroImage ,
vol. 23, pp. S208–S219, 2004.
[49] Y . Shmulev, M. Belyaev, A. D. N. Initiative et al. , “Predicting Con-
version of Mild Cognitive Impairments to Alzheimer’s Disease and
Exploring Impact of Neuroimaging,” in Graphs in Biomedical Image
Analysis and Integrating Medical Imaging and Non-Imaging Modalities .
Springer, 2018, pp. 83–91.
[50] C. Cortes and V . Vapnik, “Support-vector networks,” Machine Learning ,
vol. 20, no. 3, pp. 273–297, 1995.
[51] R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and
D. Batra, “Grad-CAM: Visual Explanations from Deep Networks via
Gradient-Based Localization,” in 2017 IEEE International Conference
on Computer Vision (ICCV) , 2017, pp. 618–626.
[52] G. Hinton and S. Roweis, “Stochastic neighbor embedding,” in Pro-
ceedings of the 15th International Conference on Neural Information
Processing Systems . MIT Press, 2002, pp. 857—-864.
Mona Ashtari-Majlan received her Master’s de-
gree in Health Systems Engineering from Amirk-
abir University of Technology, Tehran, in 2021.
She is a PhD candidate in computer science
at Universitat Oberta de Catalunya, Spain. Her
area of interest includes Biomedical Image Pro-
cessing, Computer Vision, and Deep Learning.
Abbas Seifi is a professor of Industrial Engi-
neering and Management Systems at Amirkabir
University of Technology in Iran. He did his
BASc and MASc in Industrial Engineering at
Sharif University of Technology. He received his
PhD in Systems Design Engineering from the
University of Waterloo in Canada and worked
there as a postdoctoral research associate for
over 2 years. His teaching and research interests
include optimisation and simulation of various
operational research problems, data driven op-
timisation, data science, machine learning and system dynamics.
Mohammad Mahdi Dehshibi received his PhD
in Computer Science in 2017 from IAU, Iran. He
is currently a research scientist at Universitat
Oberta de Catalunya, Spain. He was also a
visiting researcher at Unconventional Computing
Lab, UWE, Bristol, UK. He has contributed to
more than 60 papers published in scientific jour-
nals and international conferences. His research
interests include Affective Computing, Medical
data processing, Deep Learning, and Unconven-
tional Computing.