Arabian Journal for Science and Engineering (2022) 47:2201–2218

[Link]

RESEARCH ARTICLE-COMPUTER ENGINEERING AND COMPUTER SCIENCE

Detecting the Stages of Alzheimer’s Disease with Pre-trained Deep

Learning Architectures
Serkan Savaş1

Received: 10 June 2021 / Accepted: 18 August 2021 / Published online: 20 September 2021
© King Fahd University of Petroleum & Minerals 2021

Abstract
Deep learning algorithms have begun to be used in medical image processing studies, especially in the last decade. MRI is
used in the diagnosis of Alzheimer’s disease, a type of dementia disease, which is the 7th among the diseases that cause death
in the world. Alzheimer’s disease has no known cure in the literature, so it is important to attempt treatment before starting the
irreversible path by diagnosing the pre-illness stages. In this study, the previous stages of Alzheimer’s disease were classified
as normal, mild cognitive impairment, and Alzheimer’s disease through brain MRIs. Different models using CNN architecture
were used to classify 2182 image objects obtained from the ADNI database. The study was presented in a very comprehensive
comparison framework, and the performances of 29 different pre-trained models on images were evaluated. The accuracy
values of each model and the precision, specificity, and sensitivity rates of each class were determined. In the study, the
EfficientNetB0 model provided the highest accuracy at the test stage with an accuracy rate of 92.98%. In the comparative
evaluation stage with the confusion matrix, the highest rates of precision, sensitivity, and specificity values of the Alzheimer’s
disease class were achieved by EfficientNetB3 (89.78%), EfficientNetB2 (94.42%), and EfficientNetB3 (97.28%) models,
respectively. The results of the study showed that among the pre-trained models, EfficientNet models achieved a high rate of
classification performance as the models with the highest performance. This study will contribute to clinical studies in early
prevention by detecting Alzheimer’s disease before it occurs.

Keywords Deep learning · Convolutional neural network · Alzheimer’s disease · Mild cognitive impairment · Magnetic
resonance image classification · Pre-trained models

1 Introduction Changes in amyloid precursor protein (APP) cleavage and

production of the APP fragment betaamyloid (Aβ) along
Alzheimer’s disease (AD) and other forms of dementia with hyperphosphorylated tau protein aggregation coalesce
ranked as the 7th leading cause of death [1]. AD is the most to cause reduction in synaptic strength, synaptic loss, and
common type of dementia covering 60%–80% of dementia neurodegeneration. Metabolic, vascular, and inflammatory
cases. Dementia is a syndrome in which there is a deterio- changes, as well as comorbid pathologies, are key compo-
ration in cognitive function beyond what might be expected nents of the disease process [3].
from normal aging. It causes damage to memory, thinking, Dementia is usually chronic or progressive. This progres-
orientation, comprehension, calculation, learning capacity, sive process may follow a path starting from the cognitively
language, and ability to distinguish but consciousness is not normal (CN) stage, resulting in mild cognitive impairment
affected. The impairment in cognitive function is commonly (MCI) stage, and AD. For being incurable neurodegenera-
accompanied and occasionally preceded, by deterioration tive disorder [4,5], it is important to detect early stages of
in emotional control, social behavior, or motivation [2]. AD before it happens. The greatest risk factors for AD are
old age, family history, and the presence of the Apolipopro-
tein e4 (ApoE4) gene in a person’s genome [6]. MCI is one of
B Serkan Savaş
the early stages of AD and causes a slight but noticeable and
serkansavas@[Link]
measurable decline in cognitive abilities, including memory
1 Department of Computer Engineering, Faculty of and thinking skills which may cause AD or another dementia.
Engineering, Çankırı Karatekin University, 18100 Çankırı,
Turkey

123
2202 Arabian Journal for Science and Engineering (2022) 47:2201–2218

Early detection of MCI and appropriate treatment methods random neural network cluster [30], and diagnosis researches
can delay the development of AD [7]. performed with a combination of sparse regression models
Due to the irreversible end of cognitive disorders, various with deep neural network (DNN) [31]. These DL algorithms
studies are continuing to detect in the progressive process. used MRIs for analyses besides PET images were used to
Current medical research has been supported by state-of- construct cascaded CNNs to learn the multi-level and multi-
the-art analysis algorithms, especially in recent years, so AD modal features [32] and to predict the final diagnosis of AD
researches. Most researches are based on brain images and and MCI [33].
different graph-based learning [8] algorithms used to identify As being widely used in computer identification,
subjects at different progression stages of AD. Not only brain autonomous vehicles, natural language processing, hand-
images used for diagnosis of disease but also different fac- written character recognition, signature verification, voice
tors such as gender, age, education [9] are being investigated and video recognition, big data [34], image processing, and
for neurocognitive network connectivity differences between medical image processing; DL approaches are necessary for
healthy aging and MCI and associations with cognitive status automated medical decision-making systems because non-
and these factors [10]. In different studies, biological fluid automated processes are more expensive, demand intensive
biomarkers such as invasive CSF and non-invasive (blood, labor and therefore subject to human-induced errors [35].
saliva, urine, and tears) biomarkers [11,12], dry biomark- In this study, performance comparisons and classification
ers such as structural and functional imaging, and ocular successes of pre-trained CNN architectures such as AlexNet
biomarkers of AD [12] have been extensively examined, and [36], ZFNet [37], ResNet 50-101-152-50V2-101V2-152V2
the results have been explained. While some research focuses [38], and VGG 16-19 [39], which are successful on the Ima-
on clinical results to investigate the relationship between pro- geNet competition (ImageNet Large Scale Visual Recogni-
granulin expression in peripheral blood and clinical diagnosis tion Challenge - ILSVRC), and some other pre-trained mod-
of AD and MCI using microarrays [13] and variant analysis els such as LeNet [40], Xception, MobileNet, MobileNetV2,
[14], methods such as machine learning (ML) and deep learn- InceptionV3, InceptionResNetV2, DenseNet 121-169-201,
ing (DL) have been used frequently in the context of artificial NASNetMobile, NASNetLarge, and EfficientNetB0-B7 from
intelligence (AI) studies [15], especially recently. Keras Application Library [41] (29 models in total) were
Since the day it started with the question “Can machines evaluated on CN, MCI, and AD with MRIs data from
think?” [16], AI [17] has become to produce high accuracy Alzheimer’s Disease Neuroimaging Initiative [42] study plat-
results for many problems. The neurocognitive researches form.
proposed different ML algorithms such as dual-tree complex
wavelet transforms, principal component analysis, linear dis-
criminant analysis, extreme learning machine [18], dynamic 2 Materials and Methods
connectivity networks learning framework [19], one-way
analysis of variance analysis on the multiscale entropy Data used in the preparation of this article were obtained from
[20], N-fold cross-validation [21], Bayesian latent time joint the Alzheimer’s Disease Neuroimaging Initiative (ADNI)
mixed-effects [22], and ensemble feature selection approach database ([Link]). For experimental tests, ADNI1
[23] for early detection and classification of AD and for mini- Complete 3Yr 1.5T data of ADNI is chosen. This dataset
mizing the mortality rate on brain magnetic resonance images contains T1-weighted sagittal MR images. There are 2182
(MRIs). The MRI scan of the brain shows the brain structures, image items with .nii format and size of 44.16 GB which is
shrinkage of the brain, vascular irregularities, and any other provided as Archived, NIFTI, MINC, and Analyze options.
structural changes that might cause cognitive dysfunction After downloading the dataset, an image pre-processing is
[24]. Besides not only MRIs used for detection and classifi- applied at three steps:
cation but also positron emission tomography (PET) images
are used with a random forest-robust support vector machine 1. Converting images from .nii format to .png format with
for the identification of MCI [25]. Python code: In this step, there are 166 frames exported
DL algorithms have begun to assist in different applica- from each .nii image, and two middle images were chosen
tions such as health services, translation services, advertising (number 83 and number 84) from each image item.
services, driverless vehicles, film suggestions, chatbot, page 2. Splitting the dataset randomly with Python code: In this
suggestions, and many more with successful classification step, the images are separated as AD, MCI, and CN sub-
and/or detection solutions [26]. Different DL approaches clusters. From these sub-clusters, train and test images are
were also used for AD recognition [27] and to predict chosen randomly with the rate of 90% and 10%, respec-
MCI-to-AD conversion such as convolutional neural net- tively.
work (CNN) [28]. Classification studies were performed with 3. Resizing and splitting train and validation images: In this
multiple cluster dense convolutional neural networks [29], step, inside the model program before training, the train

123
Arabian Journal for Science and Engineering (2022) 47:2201–2218 2203

ADNI
DATABASE
IMAGE ACQUISITION 2182 Image Item
IMAGE EXTRACTION

4364 MR Images
DATA CLEANING

4306 MR Images
TRAIN 90% TEST 10%

430 MR
3876 MR Images
Images
TRAIN 90% VALIDATION 10%

386 MR
3490 MR Images
Images

Fig. 1 Image pre-processing steps

Table 1 Summary of the dataset was separated as data that the models would not see during
Type CN MCI AD Total training. Of the remaining 3876 images, 3490 (90%) images
were used for training and 386 (10%) images were used for
Number of patients 135 148 99 382 validation. Sample MRIs belonging to all three classes are
Male patients 66 105 52 223 shown in Fig. 2.
Female patients 69 43 47 159 To analyze the data used in the study, DNN architectures
Mean of age 77.24 75.63 75.81 76.23 created with the CNN algorithm were used. In the structure
Mean of visits 4.20 4.30 3.32 3.94 of the CNN algorithm, there are various layers and functions
Number of images 1476 1924 906 4306 such as convolution layers, pooling layers, fully connected
Train images 1196 1560 734 3490 layers, activation functions, and dropout layers. In models
Validation images 132 172 82 386 designed using this algorithm, many parameters such as the
Test images 148 192 90 430 number of layers, activation functions, dropout rate, and loca-
tions, the number of epochs, batch size, learning rate, and
optimization method are determined by the programmer who
created the model. Researchers can create their models or
images are split into train and validation images randomly work on the data using the weights of previously created
with the rate of 90% and 10%, respectively. Besides, all (pre-trained) models with the transfer learning method. In
images are resized to 224 × 224 resolution. real-world applications, it is hard or sometimes impossible to
achieve or re-collect enough data for training a CNN model.
The processes of acquiring images, image extraction, data Also, overfitting and convergence issues are the potential
cleaning, and splitting the images into train, validation, and problems that may be faced with for training a deep CNN.
test datasets are shown in Fig. 1. For this reason, the transfer-learning approach should be pre-
The summary of the dataset is shown in Table 1 after ferred to solve these kinds of problems [43]. A sample CNN
applying image pre-processing steps to the dataset. structure is shown in Fig. 3.
It is seen in Table 1 that there is a total of 382 patients con- In the convolutional layer, each unit is connected to local
sisting of 223 male and 159 female patients in the dataset. The patches in the feature maps of the previous layer through a set
number of CN, MCI, and AD labeled patients is 135, 148, of weights called a filter bank; thus, an organized feature map
and 99, respectively. The mean age of patients is 76.2. Data units are created. All units in a feature map share the same
cleaning was performed on 4364 MR images obtained after filter bank. Different feature maps in a layer use different
the image extraction process was performed, and 58 inappro- filter banks [44]. The convolution of two functions ( f ∗ g)
priate images were cleared from the database. As a result, in the finite range [0, t] is defined as Eq. (1) [45].
experiments were carried out on 4306 images. 430 (10%)
of these images were randomly selected as test data, which

123
2204 Arabian Journal for Science and Engineering (2022) 47:2201–2218

Fig. 2 Sample MRIs of a AD, b CN, c MCI classes

Fig. 3 A sample CNN structure

 t
[ f ∗ g](t) ≡ f (τ )g(t − τ )dτ (1) (ReLU) (Eq. 5).
0
1
Sigmoid f (x) = (3)
1 + ex
In Eq. (1), [ f ∗ g](t) means the convolution of the functions
2
f and g. Convolution is taken in an infinite range mostly as T an H tanh(x) = −1 (4)
calculated Eq. (2) [45]: 1 + e−2x
0 f or x<0
ReLU f (x) = (5)
 ∞ x f or x≥0
[ f ∗ g] ≡ f (τ )g(t − τ )dτ
−∞ Another activation function called softmax is widely used
 ∞
in the last layer of CNNs. The standard softmax function:
= g(τ ) f (t − τ )dτ (2)
−∞ σ : R K → [0, 1] K is defined by Eq. (6) [46,47]:

ezi
Although kernel filter was defined as 11 × 11 in AlexNet σ (z)i =  K f or i = 1, ..., K and
zj
and 7 × 7 in ZFNet, nowadays mostly 3 × 3 and 5 × 5 j=1 e

filters are being used in the literature in convolution lay- z = (z 1 , ..., z K )ER K (6)
ers. Angular edges, dark/light transition forms are evaluated
and calculated separately as features with different edge and applies the standard exponential function to each element
detection filters. Here mathematically, the filtering opera- z i of the input vector z and normalizes these values by divid-
tion performed by a feature map is a discrete convolution, ing by the sum of all these exponentials. The normalization
which is giving the name to the algorithm. The result of this ensures that the sum of the components of the output vector
locally weighted sum then passes through a nonlinearity such σ (z) is 1. These nonlinear (activation) functions are used to
as a sigmoid (Eq. 3), TanH (Eq. 4), and rectified linear unit control the output value of a neuron to decide whether a neu-

123
Arabian Journal for Science and Engineering (2022) 47:2201–2218 2205

ron will be active or not. In this respect, activation functions other problems. For this reason, studies conducted with DL
are an important feature for DNNs. models, especially in recent years, have gradually evolved
The purpose of the pooling layer is to reduce the image toward studies conducted on pre-trained models. These stud-
size by processing with an n × n matrix filter to reduce the ies either directly use the weights of pre-trained models
computational cost. It is not a mandatory layer for the CNN or produce solutions with additional layers by using these
algorithm, and some models may not use this layer. In this weights with the transfer learning method. Pre-trained CNN
layer, a filter with the determined size is passed over the image models were used in the study. The architectural structure
according to the determined number of strides. As a result of and parameter details of the models are given in Table 2.
this process, a reduced new matrix occurs. Maximum pooling The hyper-parameters to be used in the algorithm in DL
and average pooling layers are the most used ones. Besides, studies are mostly determined by trial and error depending
there are other pooling methods such as mixed pooling, LP on the intuition of the programmer, previous studies, the
pooling, stochastic pooling, and spatial pyramid pooling [48]. structure of the algorithm, the type and size of data, and sim-
Overfitting is an important problem for ML and DL. The ilar reasons. The hyper-parameters used in this study were
dropout technique reduces overfitting, thus improving the determined as follows, by trying many different alternatives
performance of NNs. Normally, learning with backpropaga- with optimum options without disrupting the structure of
tion can adapt the model to the training data, but this learning the pre-trained models. In all models, a standard structure
may not be generalized to test data. Random dropouts inside was created. Since the final layers of the selected pre-trained
the model can break these adaptations and improve the test models were designed according to 1000 classifications
results. in ImageNet competition, the output data were gradually
Dropout technique is general for different applications and reduced to classify the images used in the study as AD, CN,
was found to improve the performance of NNs in a wide vari- and MCI, and Dense layers were added to 512 and 3, respec-
ety of application domains such as object classification, digit tively. Dropout layers were added between these layers to
recognition, speech recognition, document classification, and prevent overfitting and its value was determined as 0.5 to
analysis of computational biology data. Feed-forward NN avoid inconsistent results. The optimizer is set as Adam with a
with dropout is described as [49]: default learning rate. Categorical cross-entropy was chosen
for the Loss parameter, and accuracy was chosen as met-
(l) rics. Training processes consisting of standard 250 epochs
r j ∼ Bernoulli( p),
were carried out for the models. This number of epochs was
ỹ (l) = r (l) ∗ y (l) , determined with repetitive tests to stop the training and opti-
(l+1) (l+1) l (l+1)
(7)
zi = wi ỹ + bi , mize the hyper-parameters before an overfitting condition
(l+1) (l+1)
occurred. In all models except NasneNet, images were pro-
yi = f (z i ). cessed with 224 × 224 × 3 image format, only NasneNet
structure was planned in 331 × 331 × 3 structure, so the data
where lE(1, . . . , L − 1) index is the hidden layers of the in this model were trained in the specified dimensions. The
network, i is hidden unit, z(l) denotes the vector of inputs final weights of the models were saved during the training
into layer l, y(l) denotes the vector of outputs from layer, and validation phase and tested on data that were reserved for
W (l) and b(l) are the weights and biases at layer l, and f is testing which the models had never seen. Thus, the reliability
an activation function, ∗ denotes an element-wise product, of the test results of the models was ensured. The framework
r (l) is a vector of independent Bernoulli random variables created for the training of models and test matches in the
each of which has probability p. study is shown in Fig. 4.
At the end of the CNN structure, usually, the fully con- To use the DL models on MRIs, Kaggle, which is a crowd-
nected (FC) layer operates on a flattened input where each sourced platform for data scientists to solve data science, ML,
input is connected to all neurons. FC layers can be used to and predictive analytics problems [52], and Google Colabo-
optimize objectives such as class scores [50]. The output of ratory (or Colab), which allows to write and execute arbitrary
the FC layer is normalized by the softmax activation func- python code for ML, data analysis, and education with com-
tion, which provides positive numbers that sum to one and is puting resources including graphic processor units (GPUs)
utilized to derive the output probabilities of the classification [53], were used. The DL models were trained using the Keras
layer that uses the calculated probabilities to assign the input Applications DL libraries with Tensorflow in the Python pro-
to one of the mutually exclusive target classes and computes gramming language on these platforms. After the operation
the loss [51]. of the models, graphics and weights were recorded and the
Although special models that offer efficient solutions for accuracy and loss parameters of the models were visualized.
different applications are frequently developed in the litera- To evaluate the performance of the models, Confusion Matri-
ture, the basis of DL lies in the adaptability of a solution to ces were also created.

123
2206 Arabian Journal for Science and Engineering (2022) 47:2201–2218

Table 2 Details of the models

N. Model Size Top-1 & Top-5 Acc Model params Depth TPoAD TDMpT

1 Xception 88 MB 0.790 0.945 22,910,480 126 72,189,227 66.67

2 VGG16 528 MB 0.713 0.901 138,357,544 23 27,561,795 62.50
3 VGG19 549 MB 0.713 0.900 143,667,240 26 32,871,491 66.67
4 ResNet50 98 MB 0.749 0.921 25,636,712 – 74,916,867 62.5
5 ResNet101 171 MB 0.764 0.928 44,707,176 – 93,935,107 83.33
6 ResNet152 232 MB 0.766 0.931 60,419,944 – 109,601,795 116.67
7 ResNet50V2 98 MB 0.760 0.930 25,613,800 – 74,901,635 70.83
8 ResNet101V2 171 MB 0.772 0.938 44,675,560 – 93,911,171 79.17
9 ResNet152V2 232 MB 0.780 0.942 60,380,648 – 109,570,179 104.17
10 InceptionV3 92 MB 0.779 0.937 23,851,784 159 47,984,803 75
11 InceptionResNetV2 215 MB 0.803 0.953 55,873,736 572 73,939,043 104.17
12 MobileNet 16 MB 0.704 0.895 4,253,864 88 28,899,139 58.33
13 MobileNetV2 14 MB 0.713 0.901 3,538,984 88 34,388,563 58.33
14 DenseNet121 33 MB 0.750 0.923 8,062,504 121 32,646,019 70.83
15 DenseNet169 57 MB 0.762 0.932 14,307,880 169 54,232,963 70.83
16 DenseNet201 80 MB 0.773 0.936 20,242,984 201 66,263,939 91.67
17 NASNetMobile 23 MB 0.744 0.919 5,326,716 – 30,727,957 83.33
18 NASNetLarge 343 MB 0.825 0.960 88,949,818 – 334,512,665 362.5
19 EfficientNetB0 29 MB – – 5,330,571 – 36,122,239 62.5
20 EfficientNetB1 31 MB – – 7,856,239 – 38,627,875 66.67
21 EfficientNetB2 36 MB – – 9,177,569 – 43,026,949 75
22 EfficientNetB3 48 MB – – 12,320,535 – 49,233,451 91.67
23 EfficientNetB4 75 MB – – 19,466,823 – 62,508,363 112.50
24 EfficientNetB5 118 MB – – 30,562,527 – 79,723,059 95.83
25 EfficientNetB6 166 MB – – 43,265,143 – 98,540,507 116.67
26 EfficientNetB7 256 MB – – 66,658,687 – 65,099,731 150
27 AlexNet 298 MB 0.633 0.846 – 8 26,042,435 66.67
28 ZFNet – 0.64 0.853 – 7 23,650,243 70.83
29 LeNet 539 MB – – – 7 47,172,099 58.33
TPoAD: Total Parameter of ADNI Data
TDMpT: Training Duration Minutes per Training

3 Results comparison stage because they have an accuracy rate under

50% during the training. Due to the low performance of these
To ensure the accuracy and reliability of the study, fivefold 11 models, accuracy and confusion matrix evaluations were
verification training was applied to all models, and data were continued on the remaining 18 models. The first situation
recorded in separate files during the training, validation, and that needs to be explained in the results of these excluded
testing stages. These results were then combined for each models is that certain model groups do not produce efficient
model, and the mean values and standard deviations of the results on brain MRIs. While the performance of the first ver-
models were calculated. The graphics of the results of the sions of the ResNet models reached certain rates, the models
models were analyzed comparatively over the average values arranged as V2 could not achieve this performance. Similarly,
of the five tests. In the study, a very comprehensive analy- all DenseNet and NASNet models could not achieve success
sis process was carried out by evaluating the performances on these images. Interesting findings have emerged for the
of 29 different models in total as seen in Table 2. Among ZFNet and Inception models. Although the ZFNet model
these 29 models, ResNet 50V2-101V2-152V2, InceptionV3, was established by making minor changes to the AlexNet
InceptionResNetV2, DenseNet 121-169-201, NASNetMo- model, it did not reach the level of performance that AlexNet
bile, NASNetLarge, and ZFNet models are excluded from the provided. Interestingly, the Inception model did not achieve

123
Arabian Journal for Science and Engineering (2022) 47:2201–2218 2207

Fig. 4 Train, test, and comparison framework

the same success on MRI, although it previously achieved B1 models showed more consistent results than the other
success in PET images, which are a different type of brain series of the EfficientNet models, the B6 model (Fig. 6c)
image. showed the most emission among the EfficientNet models.
After removing these models from the test and comparison The loss function is an important indicator because it is
set, the results obtained with the remaining models continue used to measure the inconsistency between the predicted
as follows. A single training graph was created for each value and the actual label. The loss layer of a NN compares
model by calculating the means and standard deviations of the output of the network with the ground truth for the case
the recorded training accuracies. The training curves of the of image processing [54]. The choice of the loss function
remaining 18 models are shown in Fig. 5, respectively. is critical for model estimation and evaluation [55]. The loss
Although almost all of the models shown in Fig. 5 per- function used in the study is determined as categorical cross-
formed a learning leap with vertical acceleration before the entropy and calculated as in Eq. 8:
first 50 epochs, the number of epochs was determined as 250
since the learning increase of other models except ResNet Out
put Si ze
models continues. This number of epochs was determined as Loss = − yi log ỹi (8)
the cutoff point before the models were overfitted the data and i=1
applied to all models. Within the training curves, the AlexNet
model achieved better accuracy training than the EfficientNet where ỹi is the i −th scalar value in the model output, yi is the
models (Fig. 5a), while the EfficientNet models decreased corresponding target value, and the output size is the number
their performance from B0 to B7 (Fig. 5a–c). While the LeNet of scalar values in the model output. In this context, yi is the
model (Fig. 5d) showed the best performance in the training probability that event i occurs and the sum of all yi is 1, mean-
graph of the models, the Xception model (Fig. 5f) showed the ing that exactly one event may occur. The minus sign ensures
lowest performance. ResNet, VGG 16-19, and MobileNet & that the loss gets smaller when the distributions get closer to
V2 models achieved similar training performance. each other. Since the Softmax is the only activation func-
Standard deviation results have also been an indicator of tion recommended to use with the categorical cross-entropy
the consistency of the models in training. Standard deviation loss function [56], it is used in these study models. Given
charts of the models are shown in Fig. 6, respectively. the importance of the loss function, it is clear that the esti-
As seen in Fig. 6, the standard deviations of all mod- mation and evaluation phases of a model are inextricably
els except the Xception model (Fig. 6f) have been realized linked. If the loss function affects the model specification,
between 0 & 0.02, proving that the models show consistency estimating a model under one loss and evaluating it under
in each training. In the Xception model, again a low standard another amount, changing the model specification without
deviation curve was observed, remaining between 0.08 & allowing the parameter estimates to be adjusted [57]. The
0.04. The models that gave the most consistent results from robustness of the model increases with decreasing value of
each training process were VGG models (Fig. 6f), ResNet the loss function, and the loss value approaching zero in ANN
models (Fig. 6e), LeNet, and MobileNet models (Fig. 6d) is specified as the desired state. In addition to the accuracy
among the models. While the AlexNet and EfficientNetB0- graphs, the loss parameters of the models were also evaluated

123
2208 Arabian Journal for Science and Engineering (2022) 47:2201–2218

Fig. 5 Train accuracy curves of the models

Fig. 6 Train accuracy standard deviation curves of the models

in the study. Loss values of the models are shown in Fig. 7, that while high accuracy was achieved in the training phase
respectively. of this model (Fig. 5d), the loss graph showed a fluctuat-
The Xception model produced the lowest performance ing course (Fig. 7d) and produced a negative result for the
results in the accuracy charts as seen in Fig. 5 and also consistency of the model accuracy. In loss charts, the lowest
produced the highest loss value in loss charts (Fig. 7f). Effi- values were obtained from ResNet models (Fig. 7e). After-
cientNetB7 model (Fig. 7c) produced the second highest loss ward, a ranking was formed as MobileNet models, VGG
value. An important case has emerged for the LeNet model models, AlexNet, and EfficientNetB0-B7 models. The loss

123
Arabian Journal for Science and Engineering (2022) 47:2201–2218 2209

Fig. 7 Loss curves of the models

Fig. 8 Standard deviations of loss curves of the models

value revealed by AlexNet also continued with oscillation in obtained in each epoch in the fivefold verification stage in a
its graph. Another remarkable result is that the loss values of similar way to the accuracy values. Standard deviation graphs
the EfficientNet models increased from B0 to B7. of loss values are shown in Fig. 8, respectively.
The models achieved success by showing low loss rates, As seen in Fig. 8, the standard deviation values of the
and these results were compared in detail during the testing loss function of the models in the study showed consistency
stage. With the loss function, the stability of the models is by remaining between 0 and 0.05 except for the two mod-
verified by evaluating the standard deviations of the values els. Among the two models, the Xception model showed

123
2210 Arabian Journal for Science and Engineering (2022) 47:2201–2218

Fig. 9 Feature extraction. a

Input Image. b 1st Layer. c 5th
Layer. d 10th Layer. e 15th
Layer

consistency by showing a 0.1 standard deviation (Fig. 8f). Testing was carried out on the recorded weights of the
However, the LeNet model produced a discrepancy with the models after the training and validation stages. The testing
standard deviation rate increasing of oscillations as the num- process is important in terms of the accuracy results that the
ber of epochs increased (Fig. 8d). To decide the performance models will produce after training. It is the most important
of this model, the results of the test stage were examined. indicator of the performance of the models on the images.
Apart from the two models, the standard deviation rates of During the testing stage, the automatic feature extraction
AlexNet, EfficientNetB0-B1 (Fig. 8a), MobileNet models process that takes place in layers of the CNN architecture
(Fig. 8d), and ResNet models (Fig. 8e) were determined as was also recorded. Sample images were also saved from the
the most consistent models. After these models, there were feature map created in layers such as the input layer, and the
VGG models (Fig. 8f) and Efficient B7-B3-B5-B4-B2-B6 1st, 5th, 10th, and 15th layers in CNN models by filtering the
models, respectively (Fig. 8b,c). image in the dataset and are shown in Fig. 9.

123
Arabian Journal for Science and Engineering (2022) 47:2201–2218 2211

Table 3 Test results of the models 0.0316 other models, this value is considered to be close to
N. Model Accuracy SD Loss zero and a successful value. The accuracy, loss, and standard
deviation graphs of the test results of the models are shown
1 AlexNet 89.95% 0.0132 0.0167 in Fig. 10.
2 EfficientNetB0 92.98% 0.0095 0.0316 Among all models, EfficientNet models achieved promis-
3 EfficientNetB1 91.91% 0.0135 0.0336 ing results with general performance conditions. It is thought
4 EfficientNetB2 90.88% 0.0076 0.0524 that the performance of these models can be further increased
5 EfficientNetB3 90.93% 0.0114 0.0393 through transfer learning methods. Thus, it has been proven
6 EfficientNetB4 88.33% 0.0217 0.0481 that the models prepared for object classification will also be
7 EfficientNetB5 89.91% 0.0145 0.0674 successful in medical image processing.
8 EfficientNetB6 89.49% 0.0088 0.0804 In the last part of the test stage, the confusion matrixes
9 EfficientNetB7 87.12% 0.0065 0.1497 of all models were created and their precision, recall (sen-
10 LeNet 87.72% 0.0114 0.1655 sitivity), and specificity values were calculated. This matrix
11 MobileNet 83.40% 0.0220 0.0208 has contributed to an in-depth study of each model’s perfor-
12 MobileNetV2 80.19% 0.0183 0.0192 mance on images of each class. These values are calculated
13 ResNet50 88.56% 0.0102 0.0145 as Eq. (9):
14 ResNet101 86.98% 0.0097 0.0289
15 ResNet152 87.77% 0.0147 0.0289
Pr ecision = T P/(T P + F P)
16 VGG16 89.72% 0.0058 0.0324
17 VGG19 89.77% 0.0082 0.0307 Recall(Sensitivit y) = T P/(T P + F N ) (9)
18 Xception 77.40% 0.0404 0.3151 Speci f icit y = T N /(T N + F P)

where TP is true positive, FP is false positive, TN is true

In the study, the results of the test process separated from negative, and FN is false negative. Precision is a measure of
the images obtained from the ADNI database and performed how accurately all classes are predicted. It is also known as
on the data that the models have never seen before are shown positive predictive value. Recall can be defined as the ratio
in Table 3. of the total number of correctly classified positive examples
When the test accuracy rates of these 18 models are exam- divide by the total number of positive examples. Specificity
ined in Table 3, the EfficientNetB0 model showed the highest indicates when it is actually negative and how often does it
performance with 92.98% accuracy on the test data. This predict negative [35].
model is followed by EfficientNetB1 (91.91%), Efficient- The confusion matrix was produced by the weights of the
NetB3 (90.93%), and EfficientNetB2 (90.88%), respectively. models after each training was recorded, and the average of
Thus, EfficientNet models verified the performance curves 5 matrixes was taken after the fivefold verification training.
in the test phase which they showed during the training and Thus, integrity has been achieved with the previous train-
produced a reliable and valid result. The models ranked at ing and testing phase. The values are not rounded to avoid
89% accuracy rate were AlexNet, EfficientNetB5, VGG19, variation in the number of images used to test the models.
VGG16, and EfficientNetB6 models, and the accuracy rates This is the justification for the floating number in the number
of these models were 89.95%, 89.91%, 89.77%, 89.72%, of images belonging to the classes. These floating sections
and 89.49%, respectively. Other models provided accuracy were also taken into account in the calculation of the values.
ranging from 88% to 77%. Among the models, the lowest The three values of each model belonging to each class are
performance rate was seen in the Xception model (77.4%). calculated and shown in Table 4.
When the standard deviations of the results obtained by Among the confusion matrix values, the models that pro-
subjecting the models to the fivefold verification test were duced the highest results for all classes were determined.
examined, it was observed that all models showed a standard Among these models in the sensitivity field, EfficientNetB2
deviation between 0 and 0.02, except for Xception (0.04), and (94.42%), EfficientNetB1 (94.35%), and EfficientNetB0
produced consistent results in each test. In the loss parameter, (94.34%) models took place, which produce very close rates
another performance indicator of the models, the lowest rate in the “AD” class. In the “CN” class, B3 (94.87%) and B1
with 0.0145 was shown by the ResNet50 model. However, (94.18%) took the first two places among the EfficientNet
the accuracy rate of this model did not support the model models, while the VGG16 model (92.91%) took the third
performance to the same extent as 88.56%. Although the place. In the “MCI” class, B0 (94.25%) and B4 (89.25%) of
EfficientNetB0 model, which has the highest accuracy rate, the EfficientNet models took the first and third places, and
produces a loss value that cannot be considered lower than the second was LeNet (90.65%).

123
2212 Arabian Journal for Science and Engineering (2022) 47:2201–2218

Fig. 10 a Accuracy, standard deviation, and b loss results of the test phase

Table 4 Confusion matrix of

Model Predicted Class Actual Recall Precision Specificity
the models
AD CN MCI (Sensitivity)

AlexNet AD 76 5.6 8.4 91.79% 84.44% 95.97%

CN 2.4 133.2 12.4 89.52% 90.00% 94.74%
MCI 4.4 10 177.6 89.52% 92.50% 93.78%
EfficientNetB0 AD 80 13.8 4.2 94.34% 81.63% 94.90%
CN 1.6 139.6 6.8 86.17% 94.32% 96.96%
MCI 3.2 8.6 180.2 94.25% 93.85% 95.22%
EfficientNetB1 AD 80,2 1,4 8,4 94.35% 89.11% 97.16%
CN 1,8 132,8 13,4 94.18% 89.73% 94.74%
MCI 3 6,8 182,2 89.31% 94.90% 95.66%
EfficientNetB2 AD 77.8 4.4 7.8 94.42% 86.44% 96.49%
CN 1.6 133.4 13 90.63% 90.14% 94.84%
MCI 3 9.4 179.6 89.62% 93.54% 94.60%
EfficientNetB3 AD 80.8 2.6 6.6 88.60% 89.78% 97.28%
CN 3.6 129.4 15 94.87% 87.43% 93.66%
MCI 6.8 4.4 180.8 89.33% 94.17% 95.08%
EfficientNetB4 AD 74.8 6.2 9 84.81% 83.11% 95.55%
CN 4.4 133.2 10.4 88.45% 90.00% 94.70%
MCI 9 11.2 171.8 89.85% 89.48% 91.54%
EfficientNetB5 AD 77.6 5.6 6.8 93.95% 86.22% 96.43%
CN 3 125.6 19.4 91.15% 84.86% 92.33%
MCI 2 6.6 183.4 87.50% 95.52% 96.10%
EfficientNetB6 AD 78.6 5 6.4 88.31% 87.33% 96.66%
CN 3.4 130.4 14.2 90.18% 88.11% 93.83%
MCI 7 9.2 175.8 89.51% 91.56% 93.07%
EfficientNetB7 AD 77.2 4.2 8.6 87.53% 85.78% 96.26%
CN 5.2 122.8 20 88.60% 82.97% 91.35%
MCI 5.8 11.6 174.6 85.93% 90.94% 92.33%

123
Arabian Journal for Science and Engineering (2022) 47:2201–2218 2213

Table 4 continued
Model Predicted Class Actual Recall Precision Specificity
AD CN MCI (Sensitivity)

LeNet AD 77 5.2 7.8 81.40% 85.56% 96.12%

CN 8.4 129.6 10 89.38% 87.57% 93.54%
MCI 9.2 10.2 172.6 90.65% 89.90% 91.90%
MobileNet AD 67.2 11.6 11.2 84.85% 74.67% 93.50%
CN 4.4 127.2 16.4 80.00% 85.95% 92.32%
MCI 7.6 20.2 164.2 85.61% 85.52% 88.33%
MobileNetV2 AD 62,2 13 14,8 77.75% 69.11% 92.06%
CN 7,8 113 27,2 81.65% 76.35% 88.00%
MCI 10 12,4 169,6 80.15% 88.33% 89.74%
ResNet50 AD 76.8 5.6 7.6 86.88% 85.33% 96.14%
CN 6.2 127.6 14.2 88.98% 86.22% 92.88%
MCI 5.4 10.2 176.4 89.00% 91.88% 93.27%
ResNet101 AD 69.6 10.2 10.2 89.23% 77.33% 94.20%
CN 2.6 135.2 10.2 83.25% 91.35% 95.22%
MCI 5.8 17 169.2 89.24% 88.13% 90.52%
ResNet152 AD 73.8 6.6 9.6 89.56% 82.00% 95.34%
CN 5.2 129.2 13.6 86.13% 87.30% 93.29%
MCI 3.4 14.2 174.4 88.26% 90.83% 92.43%
VGG16 AD 77.4 3.2 9.4 91.71% 86.00% 96.35%
CN 4 125.8 18.2 92.91% 85.00% 92.46%
MCI 3 6.4 182.6 86.87% 95.10% 95.72%
VGG19 AD 72.6 5 12.4 93.80% 80.67% 95.06%
CN 1.2 133.2 13.2 90.98% 90.24% 94.92%
MCI 3.6 8.2 180.2 87.56% 93.85% 94.73%
Xception AD 52.2 18.4 19.4 82.86% 58.00% 89.70%
CN 4 110.6 33.4 76.70% 74.73% 86.91%
MCI 6.8 15.2 170 76.30% 88.54% 89.38%

In the precision field, the first three models in the In light of all these results, it is clear that EfficientNet
“AD” class were EfficientNetB3 (89.78%), EfficientNetB1 models perform better on brain MRIs than other pre-trained
(89.11%), and EfficientNetB6 (87.33%) models, which again models. In itself, although it shows different rates between
produce very close rates. In the “CN” class, EfficientNetB0 classes, there is not much difference between them. Besides,
(94.32%) took first place, while ResNet101 took second place VGG models also produced locally successful results on
with 91.35%, and the VGG19 model took third place with certain classes of images. Therefore, applying EfficientNet
90.24%. In the “MCI” class, B5 (95.52%) and B1 (94.90%) models directly or through transfer learning in future studies
of the EfficientNet models took the first and third places, and will produce promising results in clinical studies.
the VGG16 (95.10%) model was found in the second place.
Finally, in the field of specificity, the first three models
in the “AD” class were EfficientNetB3 (97.28%), Effi-
cientNetB1 (97.16%), and EfficientNetB6 (96.66%) models, 4 Discussion
which again produce very close rates. In the “CN” class,
EfficientNetB0 (96.96%) took first place, while ResNet101 Due to the acceleration of ML and DL studies in the last
was the second with 95.22%, and the VGG19 model was decade, these techniques have been emphasized in medical
third with 94.92%. In the “MCI” class, B5 (96.10%) and B1 image processing studies as in many other fields. Particu-
(95.66%) of the EfficientNet models took the first and third larly in recent years, segmentation and classification studies
places, and the second was the VGG16 (95.72%) model. have been carried out using DL techniques. Although medical
image processing mainly covers segmentation studies, clas-
sification studies have also become widespread. The success

123
 2214

123
Table 5 Alzheimer’s disease classification studies
Author & Study Train/Test Analysis Results Data I-Type

Lu et al. [25] RF-SVM, ROC RF-RSVM: 92.16% Acc, 88.46% 120 MCI, 152 CN PET
LapSVM, kNN- Se, 96% Sp. LapSVM: 70% Acc,
SVM 80% Se, 60% Sp. kNN-SVM:
76.47% Acc, 65.38% Se, 88% Sp.
Luo et al. [27] CNN CM 93% Sp 47 AD, 34 CN MRI
Suk et al. [31] CNN/ DeepESR- CM MOLR+DeepESRNet: AD vs NC: 186 AD, 393 MCI, 226 CN MRI
Net 90.28% Acc, MC vs NC: 74.20%
Acc, JLLR+DeepESRNet: AD vs
NC: 91.02% Acc, MCI vs NC:
73.02% Acc
Ding et al. [33] CNN/InceptionV3 ROC 82% Sp 100% Se 2149 PET
Li and Liu [29] K-Means & ROC AD vs CN: 89.5% Acc 92.4% AUC. 199 AD, 403 MCI, 229 CN MRI
DenseNet MCI vs CN: 73.8% Acc 77.5%
AUC.
Jha et al. [18] DCWT, PCA, CM 90.26% Acc, 90.20% Sp, 90.27% Se 172 ADNI, 95 OASIS MRI
LDA, ELM on ADNI. 95.72% Acc, 96.59% Sp,
93.03% Se on OASIS.
Bi et al. [30] Random NN Elman NN 92.31% Acc 61 MRI
Lin et al. [28] CNN/ELM ROC 79.9% Acc, 86.1% AUC 188 AD, 401 MCI, 229 CN MRI
Jie et al. [19] FCN ROC lMCI vs eMCI: 78.8% eMCI vs CN: 99 MCI, 50 CN MRI
78.3% AUC: 78.3% and 77.1%
Liu et al. [32] CNN ROC AD vs NC: 93.26% Acc, pMCI vs 93 AD, 204 MCI, 100 CN MRI+PET
NC: 82.95% Acc
Salvatore et al. SVM Acc, Se, Sp 85% Acc, 83% Se, 87% Sp 50 AD, 50 MCI, 50 CN MRI
(2018) [15]
Buyrukoğlu [23] RF & EFS CM & AUC 91% Acc 188 AD, 402 MCI, 229 CN MRI
CNN: Convolutional Neural Network, ROC: Receiver Operating Characteristic Curves, Sp: Specificity,
Se: Sensitivity, Acc: Accuracy, AUC: Area Under the Curve, CM: Confusion Matrix, NN: Neural Network,
EFS: Ensemble Feature Selection.
Arabian Journal for Science and Engineering (2022) 47:2201–2218
Arabian Journal for Science and Engineering (2022) 47:2201–2218 2215

of AI algorithms in this regard is also seen as the biggest and situation-specific models, common models should be
factor. Classification studies about Alzheimer’s disease are created and transferred to future studies and used here. Espe-
shown in Table 5. cially considering the small number of images in the field of
As seen in Table 5, studies aimed at classifying the stages medicine and the difficulty of image acquisition and process-
of Alzheimer’s disease have gained momentum, especially ing, it is important to identify models that give common and
in recent years. While different ML and DL techniques reliable results in different imaging techniques.
are applied in these studies, the results of the studies vary The study has some limitations. The first and most impor-
between 70% and 95%. For example, the CNN algorithm tant of these is the need for a large number of images in DL
achieved 83% and 93% performance in one study [32], while studies, and in addition, the difficulty of obtaining images
the same algorithm achieved performance results varying in the field of biomedical image processing. Moreover, the
between 73% and 91% in another study [31]. While SVM, ethical approval process in the field of individual image
one of the ML algorithms, performed 85% in one study [15], acquisition in the field of biomedical image processing is
it gave results varying between 70% and 92% with hybrid also long. For this reason, a global database was used in the
methods in another study [25]. Another ML algorithm RF study. However, biomedical images in these databases are
has reached 91%–92% rates with hybrid methods in dif- often not sufficient for DL studies. Another limitation of the
ferent studies [23,25]. While InceptionV3, a DL algorithm, study is the hardware requirement in DL studies. Tools that
achieved sensitivity and specificity rates between 82% and allow the use of online ML and DL algorithms are often either
100% on PET images [33], in this study, it was excluded time-limited or there are problems such as disconnections
from the evaluation because it was below 50% accuracy dur- during use. Therefore, many tests need to be done repeat-
ing training. Except for one of the studies shown in Table 5, edly. Finally, it is the problem of viewing images presented
all of the others worked on less than 1000 images, and most of in formats such as NIFTI and DICOM by opening them for
them worked on a small number of images. Moreover, some those who work outside the medical field and do not have
of these images (10%–20%) were reserved for the validation certain programs. For this, a solution has been produced by
phase. In some studies, test images were not included in the writing a separate program that opens the image sequence
training, and in some, validation results during training were and takes the middle two images.
given as a result due to a low number of images.
In this study, the accuracy rates of varying between 90%
and 93% on the MRIs that the EfficientNet models have never 5 Conclusion
seen, and positive results between 90% and 97% in sensi-
tivity, specificity, and precision values have been indicative This study is important in terms of its scope, comparative
of the usability of these models in future clinical studies. framework, reliability, and consistency of its results. It has
Nowadays, the importance of common valid models that can been presented as an exemplary study for the use of DL algo-
be adapted to more than one dataset, instead of a single rithms, which are becoming increasingly common in every
model suitable for a single dataset, has increased. For this field, in the field of medical image processing. In the study,
reason, it has become important to transfer models that have unlike previous similar studies, not only on a single model
proven their success in different classification competitions or a few models but a comparative and comprehensive anal-
to medical image processing. With the very comprehensive ysis process was carried out on all models that have proven
and comparative analysis carried out in this study, steps have their success in different categories in international important
been taken to prevent the diversity and difference shown in competitions and on all models offered to use in DL libraries.
Table 5. The study has guiding and inspiring contributions to
Since the process of learning from the representation of both creating decision support systems in medical research
data takes place in DL techniques, studies conducted on more and transfers learning research in the field of DL. Human-
data give more consistent and accurate results. In this study, intensive studies are carried out to make quick decisions in
more images were studied in terms of consistency of the clinical studies. The lack of expertise causes the processes to
results, and at the same time, a real prediction process was be prolonged. In clinical research, getting fast results with the
carried out by not using the test data by the model during success of DL models, which are one of the sub-disciplines
the training. A step was also taken to avoid inconsisten- of AI, will both facilitate the work of experts and pave the
cies in different methods used in the literature. Considering way for getting more accurate results. Besides, a study was
that DL studies have evolved into transfer learning today, carried out in the field of DL, which will contribute to pre-
it is important to determine the most dominant model and venting the emergence of different and data-specific models
models to be used in biomedical image processing and clas- every day, and foreseeing the dissemination of models from
sification processes among models that have proven their specific to general. Because in the DL discipline, the main
success in different image classifications. Instead of subject- purpose is to establish common models that can give reliable

123
2216 Arabian Journal for Science and Engineering (2022) 47:2201–2218

results in different datasets, not just a single model suitable References

for a single data.
Another contribution of the study to the field was the 1. WHO. The top 10 causes of death (2020). [Link]
news-room/fact-sheets/detail/the-top-10-causes-of-death.
determination of the use of DL models in biomedical image
2. WHO. Dementia (2020). [Link]
classification as well as in object identification. Thus, by sheets/detail/dementia.
avoiding the confusion of different models in biomedical 3. Soria Lopez, J.A.; González, H.M.; Léger, Gabriel C.: Chapter 13
image processing, a guiding study was carried out in new - alzheimer’s disease. In Steven T. Dekosky and Sanjay Asthana,
editors, Geriatric Neurology, volume 167 of Handbook of Clini-
studies. Similar studies can be tested with different image
cal Neurology, pp. 231–255. Elsevier, (2019). [Link]
sequences such as T1-w, T2-w, and FLAIR, as well as with 1016/B978-0-12-804766-8.00013-3. [Link]
different planes such as coronal and axial, and the perfor- com/science/article/pii/B9780128047668000133.
mances of the models can be compared. Thus, comparative 4. Penney, J.; Ralvenius, W.T.; Tsai, L.-H.: Modeling Alzheimer’s
disease with ipsc-derived brain cells. Mol. Psych. 25(1), 148–167
options will be presented to researchers for early diagnosis
(2020). [Link]
of the disease. Because of the increasing number of different 5. Wan, Y.-W.; Al-Ouran, R.; Mangleburg, C.G.; Perumal, T.M.; Lee,
types of biomedical images, the need for standard models that T.V.; Allison, K.; Swarup, V.; Funk, C.C.; Gaiteri, C.; Allen, M.;
produce common solutions has arisen. This study is presented Wang, M.; Neuner, S.M.; Kaczorowski, C.C.; Philip, V.M.; How-
ell, G.R.; Martini-Stoica, H.; Zheng, H.; Mei, H.; Zhong, X.;
not only as a comparative study but also as a comprehensive Kim, J.W.; Dawson, V.L.; Dawson, T.M.; Pao, P.-C.; Tsai, L.-H.;
classification study in terms of scope. Classification with high Haure-Mirande, J.-V.; Ehrlich, M.E.; Chakrabarty, P.; Levites, Y.;
accuracy rates has been carried out to determine the stages Wang, X.; Dammer, E.B.; Srivastava, G.; Mukherjee, S.; Sieberts,
of Alzheimer’s disease, which is important enough to be in S.K.; Omberg, L.; Dang, K.D.; Eddy, J.A.; Snyder, P.; Chae, Y.;
Amberkar, S.; Wei, W.; Hide, W.; Preuss, C.; Ergun, A.; Ebert,
the top 10 in the list of fatal diseases in the world and to take P.J.; Airey, D.C.; Mostafavi, S.; Yu, L.; Klein, H.-U.; Carter, G.W.;
precautions before reaching the point that cannot be reversed. Collier, D.A.; Golde, T.E.; Levey, A.I.; Bennett, D.A.; Estrada,
K.; Matthew Townsend, T.; Zhang, B.; Schadt, E.; De Jager, P.L.;
Acknowledgements Data collection and sharing for this project were Price, N.D.; Ertekin-Taner, N.; Liu, Z.; Shulman, J.M.; Mangravite,
funded by the Alzheimer’s Disease Neuroimaging Initiative (ADNI) L.M.; Logsdon, B.A.: Meta-analysis of the alzheimer’s disease
(National Institutes of Health Grant U01 AG024904) and DOD ADNI human brain transcriptome and functional dissection in mouse
(Department of Defense award number W81XWH-12-2-0012). ADNI models. Cell Rep. 32(2):107908, (2020). ISSN 2211-1247. https://
is funded by the National Institute on Aging, the National Institute [Link]/10.1016/[Link].2020.107908. [Link]
of Biomedical Imaging and Bioengineering, and through generous com/science/article/pii/S2211124720308895.
contributions from the following: AbbVie, Alzheimer’s Association; 6. Leandrou, S.; Petroudi, S.; Kyriacou, P.A.; Reyes-Aldasoro, C.C.;
Alzheimer’s Drug Discovery Foundation; Araclon Biotech; BioClin- Pattichis, C.S.: Quantitative MRI brain studies in mild cognitive
ica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.; impairment and Alzheimer’s disease: a methodological review.
Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Com- IEEE Rev. Biomed. Eng. 11, 97–111 (2018). [Link]
pany; EuroImmun; F. Hoffmann-La Roche Ltd and its affiliated com- 1109/rbme.2018.2796598.
pany Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen 7. Alzheimer’s Association.: 2015 Alzheimer’s disease facts and fig-
Alzheimer Immunotherapy Research & Development, LLC.; Johnson ures. Alzheimer’s and Dementia 11(3):332–384, (2015). https://
& Johnson Pharmaceutical Research & Development LLC.; Lumosity; [Link]/10.1016/[Link].2015.02.003.
Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; Neu- 8. Wang, Z.; Zhu, X.; Adeli, E.; Zhu, Y.; Nie, F.; Munsell, B.;
roRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Guorong, W.: Multi-modal classification of neurodegenerative dis-
Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceu- ease by progressive graph-based transductive learning. Med. Image
tical Company; and Transition Therapeutics. The Canadian Institutes Anal. 39, 218–230 (2017). [Link]
of Health Research is providing funds to support ADNI clinical sites 05.003.
in Canada. Private sector contributions are facilitated by the Founda- 9. Hadjichrysanthou, C.; McRae-McKee, K.; Evans, S.; de Wolf, F.;
tion for the National Institutes of Health ([Link]). The grantee Anderson, R.M.: Initiative for the Alzheimer’s Disease Neuroimag-
organization is the Northern California Institute for Research and Edu- ing: Potential factors associated with cognitive improvement of
cation, and the study is coordinated by the Alzheimer’s Therapeutic individuals diagnosed with mild cognitive impairment or dementia
Research Institute at the University of Southern California. ADNI data in longitudinal studies. J. Alzheimer’s Dis. 66, 587–600 (2018).
are disseminated by the Laboratory for Neuro Imaging at the University [Link]
of Southern California. 10. Sullivan, M.D.; Anderson, J.A.E.; Turner, G.R.; Nathan Spreng, R.:
Intrinsic neurocognitive network connectivity differences between
Funding There is no funding for this research. normal aging and mild cognitive impairment are associated with
cognitive status and age. Neurobiol. Aging 73:219–228 (2019).
[Link]
Declarations 11. Molinuevo, J.L.; Ayton, S.; Batrla, R.; Bednar, M.M.; Bittner,
T.; Cummings, J.; Fagan, A.M.; Hampel, H.; Mielke, M.M.;
Mikulskis, A.; O’Bryant, S.; Scheltens, P.; Sevigny, J.; Shaw,
Conflicts of interest The author declares that there is no conflict of L.M.; Soares, H.D.; Tong, G.; Trojanowski, J.Q.; Zetterberg, H.;
interest. Blennow, K.: Current state of Alzheimer’s fluid biomarkers. Acta
Neuropathol. 136(6), 821–853 (2018). [Link]
s00401-018-1932-x.

123
Arabian Journal for Science and Engineering (2022) 47:2201–2218 2217

12. Villa, C.; Lavitrano, M.; Salvatore, E.; Combi, R.: Molecular and 26. Savaş, S.; Topaloğlu, N.; Kazcı, Ö.; Koşar, P.N.: Performance
imaging biomarkers in Alzheimer’s disease: A focus on recent comparison of carotid artery intima media thickness classification
insights. J. Pers. Med. 10(3), 1 (2020). [Link] by deep learning methods. In SETSCI Conference Proceedings,
jpm10030061. volume 4(5), pp. 125–131. International Congress on Human–
13. Cooper, Yonatan A.; Nachun, Daniel; Dokuru, Deepika; Yang, Computer Interaction, Optimization, and Robotic Applications
Zhongan; Karydas, Anna M.; Serrero, Ginette; Yue, Binbin; (2019). [Link]
Initiative, Alzheimer’s Disease Neuroimaging.; Boxer, Adam 27. Luo, S.; Li, X.L.J.: Automatic Alzheimer’s disease recognition
L.; Miller, Bruce L.; Coppola, Giovanni: Progranulin levels in from mri data using deep learning method. J. Appl. Math. Phys.
blood in alzheimer’s disease and mild cognitive impairment. 5(9), 1892–1898 (2017)
Annals of Clinical and Translational Neurology 5(5), 616–629 28. Lin, W.; Tong, T.; Gao, Q.; Guo, D.; Du, X.; Yang, Y.; Guo, G.;
(2018) ISSN 2328-9503. [Link] https:// Xiao, M.; Du, M.; Qu, X.: The Alzheimer’s Disease Neuroimaging
[Link]/doi/abs/10.1002/acn3.560. Initiative. Convolutional neural networks-based MRI image anal-
14. Goldstein, F.C.; Milloy, A.; Loring, D.W.: Incremental validity ysis for the Alzheimer’s disease prediction from mild cognitive
of montreal cognitive assessment index scores in mild cognitive impairment. Front. Neurosci. 12(777), 1 (2018). [Link]
impairment and alzheimer disease. Dement. Geriatr. Cogn. Dis- 10.3389/fnins.2018.00777.
ord. 45(1), 49–55 (2018). [Link] 29. Li, F.; Liu, M.: Alzheimer’s disease diagnosis based on multi-
15. Salvatore, C.; Cerasa, A.; Castiglioni, I.: Mri characterizes the ple cluster dense convolutional networks. Comput. Med. Imag.
progressive course of ad and predicts conversion to Alzheimer’s Gr. 70, 101–110 (2018). [Link]
dementia 24 months before probable diagnosis. Front. Aging 2018.09.009.
Neurosci. 10(135), 1 (2018). [Link] 30. Bi, X.; Jiang, Q.; Sun, Q.; Shu, Q.; Liu, Y.: Analysis of Alzheimer’s
00135. disease based on the random neural network cluster in FMRI.
16. Turing, A.M.: I. Computing machinery and intelligence. Mind Front. Neuroinf. 12(60), 1 (2018). [Link]
59(236), 433–460 (1950). [Link] 2018.00060.
433. 31. Suk, H.-I.; Lee, S.-W.; Shen, D.: Deep ensemble learning of sparse
17. McCarthy, J.; Minsky, M.L.; Rochester, N.; Shannon, C.E.: A regression models for brain disease diagnosis. Med. Image Anal.
proposal for the Dartmouth summer conference on artificial intel- 37, 101–113 (2017). [Link]
ligence. Rep. Dartmouth Coll. 31(08), 1955 (1955) 32. Liu, M.; Cheng, D.; Wang, K.; Wang, Y.: Initiative the Alzheimer’s
18. Jha, D.; Alam, S.; Pyun, J.-Y.; Lee, K.; Ho, K., Goo-Rak.: Disease Neuroimaging: Multi-modality cascaded convolutional
Alzheimer’s disease detection using extreme learning machine, neural networks for Alzheimer’s disease diagnosis. Neuroinfor-
complex dual tree wavelet principal coefficients and linear dis- matics 16(3):295–308. [Link]
criminant analysis. J. Med. Imag. Health Inf. 8(5), 881–890 (2018). 4.
[Link] 33. Yiming, D.; Jae, H.S.; Kawczynski, M.G.; Trivedi, H.; Harnish,
19. Jie, B.; Liu, M.; Shen, D.: Integration of temporal and spatial prop- R.; Jenkins, N.W.; Lituiev, D.; Copeland, T.P.; Aboian, M.S.;
erties of dynamic connectivity networks for automatic diagnosis of Aparici, C.M.; Behr, S.C.; Flavell, R.R.; Huang, S.-Y.; Zalo-
brain disease. Med. Image Anal. 47, 81–94 (2018). [Link] cusky, K.A.; Nardo, L.; Seo, Y.; Hawkins, R.A.; Pampaloni, M.H.;
10.1016/[Link].2018.03.013. Hadley, D.; Franc, B.L.: A deep learning model to predict a diag-
20. Niu, Yan; Wang, Bin; Zhou, Mengni; Xue, Jiayue; Shapour, Habib; nosis of Alzheimer disease by using 18f-fdg pet of the brain.
Cao, Rui; Cui, Xiaohong; Jinglong, Wu.; Xiang, Jie: Dynamic Radiology 290(2), 456–464 (2019). [Link]
complexity of spontaneous bold activity in alzheimer’s disease 2018180958.
and mild cognitive impairment using multiscale entropy analysis. 34. Yapıcı, M.M.; Tekerek, A.; Topaloğlu, N.: Literature review of
Front. Neurosci. 12(677), 1 (2018). [Link] deep learning research areas. Gazi Mühendislik Bilimleri Dergisi
2018.00677. (GMBD) 5(3), 188–215 (2019)
21. Sampath, R.; Indumathi, J.: Earlier detection of Alzheimer disease 35. Savaş, S.; Topaloğlu, N.; Kazcı, Ö.; Koşar, P.N.: Classification of
using n-fold cross validation approach. J. Med. Syst. 42(11), 217 carotid artery intima media thickness ultrasound images with deep
(2018). [Link] learning. J. Med. Syst. 43(8), 273 (2019). [Link]
22. Li, D.; Iddi, S.; Thompson, W.K.; Rafii, M.S.; Aisen, P.S.; s10916-019-1406-2.
Donohue, M.C.; Initiative, Alzheimer’s Disease Neuroimaging.: 36. Krizhevsky, A.; Sutskever, I.; Hinton, G.E.: Imagenet classification
Bayesian latent time joint mixed-effects model of progression in with deep convolutional neural networks. Adv. Neural. Inf. Process.
the Alzheimer’s disease neuroimaging initiative. Alzheimer’s and Syst. 25, 1097–1105 (2012)
Dement. Diagnos., Assessm. Dis. Monit. 10(1), 657–668 (2018). 37. Zeiler, M.D.; Fergus, R.: Visualizing and understanding convolu-
[Link] tional networks. In: European conference on computer vision, pp.
23. Buyrukoğlu, S.: Early detection of alzheimer’s disease using data 818–833. Springer (2014).
mining: comparision of ensemble feature selection approaches. 38. He, K.; Zhang, X.; Ren, S.; Sun, J.: Deep residual learning for image
Konya Mühendislik Bilimleri Dergisi 9(1), 50–61 (2021). https:// recognition. In: Proceedings of the IEEE conference on computer
[Link]/10.36306/konjes.731624. vision and pattern recognition, pp. 770–778 (2016).
24. Rajendra Acharya, U.; Fernandes, S.L.; WeiKoh, J.E.; Ciaccio, 39. Simonyan, K.; Zisserman, A.: Very deep convolutional networks
E.J.; Mohd, F.; Mohd, K.; John, T.U.; Rajinikanth, V.; Yeong, for large-scale image recognition. arXiv preprint arXiv:1409.1556
C.H.: Automated detection of alzheimer’s disease using brain (2014).
mri images-a study with various feature extraction techniques. 40. LeCun, Y.; Bottou, L.; Bengio, Y.; Haffner, P.: Gradient-based
J. Med. Syst. 43(9), 302 (2019). [Link] learning applied to document recognition. Proc. IEEE 86(11),
019-1428-9. 2278–2324 (1998)
25. Lu, S.; Xia, Y.; Cai, W.; Fulham, M.; Feng, D.D.: Early identi- 41. Keras. Keras applications (2021). [Link]
fication of mild cognitive impairment using incomplete random applications/.
forest-robust support vector machine and fdg-pet imaging. Com- 42. ADNI. Alzheimer’s disease neuroimaging initiative (2021). http://
put. Med. Imag. Gr. 60, 35–41 (2017). [Link] [Link]/.
compmedimag.2017.01.001.

123
2218 Arabian Journal for Science and Engineering (2022) 47:2201–2218

43. Daldal, N.; Cömert, Z.; Polat, K.: Automatic determination of dig- 51. Toprak, T.; Selver, M.A.; Secmen, M.; Zoral, E.Y.: Utilizing reso-
ital modulation types with different noises using convolutional nant scattering signal characteristics via deep learning for improved
neural network based on time-frequency information. Appl. Soft classification of complex targets. Turk. J. Electr. Eng. Comput. Sci.
Comput. 86, 105834 (2020). [Link] 29(1), 334–348 (2021). [Link]
105834 52. Usmani, Z.: (2018). [Link]
44. LeCun, Y.; Bengio, Y.; Hinton, G.: Deep learning. Nature 44916.
521(7553), 436–444 (2015) 53. Colab.: (2021). [Link]
45. Weisstein, E.W.: Convolution (2021). [Link] html.
com/[Link]. 54. Zhao, H.; Gallo, O.; Frosio, I.; Kautz, J.: Loss functions for image
46. Gao, B.; Pavel, L.: On the properties of the softmax function restoration with neural networks. IEEE Trans. Comput. Imaging
with application in game theory and reinforcement learning. arXiv 3(1), 47–57 (2017). [Link]
preprint arXiv:1704.00805 (2017). 55. Engle, R.F.: On the limitations of comparing mean square forecast
47. Wikipedia. Softmax function, (2021). [Link] errors: Comment. J. Forecast. 12(8), 642–644 (1993). [Link]
wiki/Softmax_function. org/10.1002/for.3980120805.
48. Gholamalinezhad, H.; Khosravi, H.: Pooling methods in deep neu- 56. Peltarion.: Categorical crossentropy math, (2021). https://
ral networks, a review. arXiv:2009.07485 (2020). [Link]/knowledge-center/documentation/modeling-view/
49. Srivastava, N.; Hinton, G.; Krizhevsky, A.; Sutskever, I.; Salakhut- build-an-ai-model/loss-functions/categorical-crossentropy.
dinov, R.: Dropout: A simple way to prevent neural networks from 57. Christoffersen, P.; Jacobs, K.: The importance of the loss function
overfitting. J. Mach. Learn. Res. 15, 1929–1958 (2014) in option valuation. J. Financ. Econ. 72(2), 291–318 (2004). https://
50. Amidi, A.; Amidi, S.: Convolutional neural networks cheat- [Link]/10.1016/[Link].2003.02.001.
sheet (2018). [Link]
cheatsheet-convolutional-neural-networks.

123