Automated Alzheimer’s Disease Diagnosis using Convolutional
Neural Networks and Magnetic Resonance Imaging
ASMAA NASR MOHAMMED1, ABDULGANI ALBAGUL2, MOAMER MUSBAH AHMAD3
1College of Electronic Technology Baniwalid
2Libyan Center for Engineering Research and Information Technology
3University of Baniwalid
Baniwalid
LIBYA
Abstract: - Alzheimer’s disease is a debilitating neuro-logical condition affecting millions globally; therefore,
correct diagnosis plays a significant role in treating or managing it effectively. Convolutional neural networks
(CNNs), which are popular deep learning algorithms are applied to image processing tasks, offer a good technique
to study and investigate images processing. In this study, a CNN model for classifying Alzheimer’s patients is
proposed. The research yielded impressive results: recall and precision scores as high as 0.9958 which indicate
trustworthy identification of true positives while maintaining few false positives; test accuracy exceeding 99%
confirming desirable generalization capabilities from the training dataset to live scenarios; ROC AUC score at an
astronomical height of 0.9999 signifying great potential in distinguishing between afflicted individuals from their
non-affected counterparts accurately. The proposed network achieved a classification accuracy of 99.94% on
LMCI vs EMCI, 99.87% on LMCI vs MCI, 99.95% on LMCI vs AD, 99.94% on LMCI vs CN, 99.99% on CN
vs AD, 99.99% on CN vs EMCI, 99.99% on CN vs MCI, 99.99% on AD vs EMCI, 99.98% on AD vs MCI,
and 99.96% on MCI vs EMCI. The proposed CNNs model is compared with two ultramodern models such as
VGG19 and ResNet50. The results show that the proposed model achieved a superior performance in diagnostic
precision and effectiveness of Alzheimer’s disease, leading to early detection, enhanced treatment plans, and
enriching the quality of life for those affected.
Key-words: - Leave Alzheimer’s disease, MRI, CNN, Transfer Learning.
Received: June 16, 2022. Revised: August 19, 2023. Accepted: September 21, 2023. Published: October 4, 2023.
1 Introduction
Alzheimer’s disease is a lasting condition that affects
countless individuals, across the globe. It stands as
the contributor to dementia accounting for 60-80% of
diagnosed cases [1]. The illness is defined by the
buildup of proteins in the brain such as beta amyloid
plaques and tau tangles. These contribute to the
depletion of neurons and synapses resulting in a
decline in abilities, over time [2]. Alzheimer’s
disease usually starts with memory loss and
confusion which then worsens into significant
cognitive difficulties. These challenges can include
issues with language, abilities, and executive
function. Sadly, there is no cure for Alzheimer’s
disease now. The treatments available only aim to
alleviate symptoms and slow down the progression of
the illness. Ongoing research is being conducted to
explore the factors behind Alzheimer’s disease and
find treatments. The main aim is to understand the
mechanisms that drive this disease and develop
therapies to address it. Promising areas of study
involve the creation of medications that specifically
target beta amyloid and tau proteins alongside
pharmacological approaches, like making lifestyle
adjustments and engaging in cognitive training [3].
Studies conducted recently have also showed a
connection between Alzheimer’s disease and other
health conditions like diabetes, high blood pressure
and depression. These findings underscore the
significance of adopting an approach towards treating
and preventing Alzheimer’s disease [4, 5, 6]. In this
research we aim to explore the effectiveness of a
CNN model in diagnosing AD using MRI scans.
Currently AD diagnosis relies on behavioural tests
which may have exabit margin for error. One of the
changes in the brains of AD patients is the atrophy of
the hippocampus and cortex well as other structural
alterations visible in MRI scans. CNN models have
shown results in using MRI images to find signs of
AD. These models could understand patterns within
pictures. The primary aim of our project is to develop
and evaluate a CNN model specifically designed for
diagnosing AD based on MRI scans. We will use a
dataset consisting of MRI scans from both AD
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.13
Asmaa Nasr Mohammed,
Abdulgani Albagul, Moamer Musbah Ahmad
E-ISSN: 2224-3488
118
Volume 19, 2023
patients and healthy individuals to train and evaluate
our model’s accuracy, sensitivity, and specificity. We
will also compare its performance with existing
techniques used for AD diagnosis. This research has
the potential to contribute towards creating an
effective technique for diagnosing AD. Timely and
correct diagnosis can lead to interventions and
improved outcomes for individuals affected by this
condition. Integrating CNN technology with MRI
scans in the diagnosis process may offer a systematic
approach towards finding cases of Alzheimer’s
disease.
2 Methodology
Study design: We conducted a retrospective study
using MRI data from patients with clinically
diagnosed AD and age-matched healthy controls.
We collected T1-weighted axial MRI scans from the
ADNI dataset and OASIS3 dataset. The MRI data
were pre-processed using the following steps.
2.1 Skull Stripping
Removing the skull, known as skull stripping, is an
initial step in neuroimaging. It involves separating
the brain tissue from the surrounding brain tissue
and skull. This process is commonly performed to
enhance the precision of image analysis techniques,
such, as segmentation and registration
[7, 8]. There are methods for skull stripping, such as
thresholding, region growing and utilizing machine
learning based approaches [9, 10]. Properly
removing the skull is extremely important in
neuroimaging applications such as functional MRI,
diffusion tensor imaging and positron emission
tomography. Skull stripping was done by examining
the image, cropping it to eliminate any surrounding
light box. Producing an image by applying a
threshold. It then gets rid of specks of noise from the
image and seals off the bottom part of the image. To
create a version the binary image undergoes erosion
before being used to mask the initial grayscale image
effectively eliminating any gaps in the binary image
as shown in Fig. 1 and Fig. 2.
2.2 Image Cropping
With precision in mind, we perform a series of steps
aimed at obtaining an optimally sized and shaped
image. Following calculations designed for scaling
purposes, our discrepancy metric reduces to a
diminutive proportion at just -10 (-2% of original
size). Through careful manipulation of crop
parameters based on user-defined bounds via bbox,
we carve out only those areas necessary for proper
viewing. The result is an accurately sized rectangle
with dimensions exactly totaling up to no more than
20 total pixels (measuring both width and height)
resulting in an image as shown in figure 3.
Fig. 1: Skull stripping stages
Fig. 2: Skull stripped image.
Fig. 3: Image cropping
2.3 Dataset Size
The MRI dataset for Alzheimer’s disease
classification consists of five classes: AD, CN,
EMCI, LMCI, and MCI. The AD class includes
individuals with Alzheimer’s disease, while the CN
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.13
Asmaa Nasr Mohammed,
Abdulgani Albagul, Moamer Musbah Ahmad
E-ISSN: 2224-3488
119
Volume 19, 2023
class includes cognitively normal individuals. The
EMCI, LMCI, and MCI classes include individuals
with mild cognitive impairment in early, late, and
progressive stages, respectively. The dataset is quite
large, with a total of 40,077 MRI images, and it
provides valuable insights into the brain changes that
occur in individuals with Alzheimer’s disease and
related conditions. Each class in the dataset has a
substantial number of MRI brain images, with 8000
pictures per class.
2.4 Convolutional Neural Network
Convolutional neural networks (CNNs) are a kind of
deep learning technique employed to analyse images
and videos. They consist of layers that execute
convolutions, pooling, and nonlinear activations.
These networks are trained using backpropagation
to enhance a loss function [11, 12]. Convolutional
Neural Networks (CNNs) have proven performance
in computer vision tasks, such as accurately
classifying images detecting objects and segmenting
visual data [13, 14, 15]. They have also been used in
fields, including natural language processing and
the recognition of speech [16, 17]. Recent
developments in CNNs include the creation of
deeper and more intricate structures, such attention
mechanisms, and residual networks, as well as the
incorporation of CNNs with other deep learning
methods, like generative adversarial networks and
reinforcement learning [18, 19]. The CNN
architecture primarily consists of five layers; Input
Layer, Convolutional Layer, Pooling Layer, Fully
Connected Layer and Output Layer [20]. These 5
CNN neural network layers as shown in fig 4.
Fig 4 Convolutional Neural Network [21]
2.5 The Proposed Architecture
Convolutional neural networks (CNNs), a type of
deep learning technology [21], have recently
demonstrated potential in the detection of
Alzheimer’s disease using MRI images. In this paper,
we suggest a CNN architecture for MRI scan-based
Alzheimer’s disease identification. A few
convolutional, pooling, and fully connected layers
make up the proposed model, which is intended to
extract and learn characteristics from the MRI
images. To avoid overfitting, the model additionally
uses regularization techniques like Dropout and L2
regularization. We think that our suggested model
will increase the reliability of MRI scans used to
identify Alzheimer’s disease and assist physicians in
making an early diagnosis when therapy is most
successful. The proposed CNN model has 29 layers,
including: 2 Conv2D layers, 2 MaxPooling2D layers,
6 SeparableConv2D layers, 6 Batch- Normalization
layers, 3 MaxPool2D layers, 5 Dropout layers, 1
Flatten layer, 4 Dense layers. The proposed
architecture is shown in Fig 5, and the proposed
structure layout is shown in Fig. 6.
Fig. 5: The proposed architecture
Fig. 6: The proposed network structural layout
2.6 Transfer Learning
In recent times people have been taking advantage of
a technique called” transfer learning” when
conducting image classification tasks due to its
incredible utility potential. Pre trained models are key
players in this technique since they allow features
extraction from images with ease. Conventionally
these models include deep neural networks trained
on vast datasets such as ImageNet; their primary goal
being able in finding varying objects and patterns
within images. Essentially transfer learning exploits
this by using the preexisting model to form a
foundation for developing recognition capabilities
for novel object categories [22]. As part of our
comparison exercise, we will be using two of the
most successful pre-trained models - VGG19 and
ResNet50 -for image classification. Their excellent
history in performing various image classification
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.13
Asmaa Nasr Mohammed,
Abdulgani Albagul, Moamer Musbah Ahmad
E-ISSN: 2224-3488
120
Volume 19, 2023
tasks has made them a preferred choice for transfer
learning applications.
2.7 Model Training
The CNN training pipeline for Alzheimer’s
detection is shown in Fig. 7.
Fig. 7: The CNN Training Pipeline Diagram
The training procedure of the CNN model to classify
images is carried out using Python and well-known
packages such as Tensor-Flow and Keras. We also
constructed our image processing in Python, utilizing
tools like OpenCV and NumPy for operations like
resizing, cropping, and normalization. To start we
resized all the images in the dataset to a size of
150x150 pixels. This helped us ensure that the input
size for the model was standardized. Additionally, we
transformed the labels from their format into a
format enabling us to employ a multi class
classification model. Afterwards we divided the
dataset into two parts. One, for training and the other
for testing. We made sure to allocate 80% of the
data for training purposes while keeping 20% aside
to assess how well the model performed on unseen
data. In our training process we utilized the entropy
loss function, a widely employed method, for solving
multi class classification problems. Additionally, we
employed the Adam optimizer, which’s an adaptive
learning rate optimization algorithm that proves to be
highly effective for learning models. To avoid
overfitting, we implemented a stopping mechanism,
with a patience of 10 epochs. Essentially this means
that if the validation loss didn’t show any
improvement, for 10 epochs the training process
would be halted prematurely to prevent the model
from fitting to the training data. To maintain the
model’s convergence, towards a solution, without any
overshooting or divergence we decided to decrease the
learning rate by a factor of 10 after each epoch. This
adjustment significantly enhanced the stability and
performance of the model during training. In the end
we made sure to train the model for 50 epochs
allowing time for it to grasp the underlying
patterns, in the data and attain performance. Through
these methods we successfully trained a learning
model that excels at classifying images, across
categories.
2.7 Evaluation Metrics
Once the training of the network (CNN) model was
complete, a range of metrics was tried to assess its
effectiveness. These metrics encompassed accuracy,
sensitivity, specificity positive predictive value
(PPV) negative predictive value (NPV) and the area
under the operating curve (AUC ROC). Accuracy is
a used measure to evaluate how well a model
performs. It quantifies the percentage of classified
samples, among the number of samples. Sensitivity
and specificity are metrics used to assess how well a
model can accurately identify negative samples.
Sensitivity gauges the ratio of identified samples to
all positive samples whereas specificity calculates the
ratio of correctly identified negative samples, to all
negative samples [23]. PPV and NPV are metrics
used to evaluate the negative values of a model. PPV
calculates the ratio of predictions out of all samples
predicted as positive while NPV calculates the ratio
of correct negative predictions, out of all samples
predicted as negative [24]. The AUC ROC is a metric
that evaluates how well a model can differentiate
between negative samples. It is calculated by plotting
the sensitivity (rate) against the specificity (1. False
positive rate), at different classification thresholds.
To evaluate how well the model performed on the
testing data we calculated the confusion matrix and
the classification report. The confusion matrix gives
us an overview of the model’s predictions in terms of
positives true negatives, false positives, and false
negatives. The classification report provides a
summary of the precision, recall and F1 score for
each class predicted by the model. We utilized these
measurements to assess how the CNN model
performed on the test data and acquire an
understanding of its merits and drawbacks. This
enabled us to pinpoint areas that needed enhancement
and implement any required modifications to the
model.
3 Results and Discussion
CNN models are commonly used for image
classification tasks, and their performance is typically
evaluated using metrics such as accuracy and loss. Ac-
curacy measures the proportion of correctly classified
samples out of the total number of samples in the
dataset, while loss measures the difference between
the predicted and actual values for each sample, with
lower values indicating better performance. During
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.13
Asmaa Nasr Mohammed,
Abdulgani Albagul, Moamer Musbah Ahmad
E-ISSN: 2224-3488
121
Volume 19, 2023
the training process, the model’s accuracy and loss are
constantly updated based on its performance on the
validation dataset. As the model learns to recognize
patterns in the images, its accuracy improves, and its
loss decreases as shown in Fig. 8 and Fig. 9.
Fig. 8: Model accuracy
Fig. 9: Model loss
The proposed CNN’s test accuracy of 99.92% is also
very high, indicating that the model is performing
well on the test set. This high accuracy suggests that
the model is generalizing well to new data and is not
overfitting to the training set. Analysis of Fig. 10
confusion matrix suggests that the proposed CNN
model is doing an exceptional job with its designated
task. The proposed CNN model has high precision
and recall for each class and f1-score of the accuracy
of 100%, additionally the proposed CNN’s recall or
sensitivity and precision scores of 99.58% and
specificity of 99.89% as shown in table 1 indicate that
it is performing very well in correctly identifying
positive cases while minimizing false positives.
Furthermore, as shown in fig 11 the model’s ROC
AUC score of 0.9999 indicates that it can effectively
distinguish between positive and negative cases, with
very few misclassifications.
Fig. 10. The proposed model’s Confusion
Table 1: The proposed model’s classification report
Fig. 11: The proposed model’s ROC curve
This is a very impressive score and suggests that the
model is highly discriminative. The PPV and NPV
for all classes are 1.00, indicating that the
classification model has a very high accuracy for both
positive and negative predictions. Our network
achieved a classification accuracy of 99.94% on
Precision
Recall
f1-score
Support
LMCI (0)
1.00
1.00
1.00
1662
EMCI (1)
0.99
1.00
1.00
1612
MCI (2)
1.00
0.99
0.99
1589
AD (3)
1.00
1.00
1.00
1542
CN (4)
1.00
1.00
1.00
1611
Accuracy
1.00
8016
Macro avg
1.00
1.00
1.00
8061
Weighted avg
1.00
1.00
1.00
8061
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.13
Asmaa Nasr Mohammed,
Abdulgani Albagul, Moamer Musbah Ahmad
E-ISSN: 2224-3488
122
Volume 19, 2023
LMCI vs EMCI, 99.87% on LMCI vs MCI, 99.95%
on LMCI vs AD, 99.94% on LMCI vs CN, 99.99%
on CN vs AD, 99.99% on CN vs EMCI, 99.99% on
CN vs MCI, 99.99% on AD vs EMCI, 99.98% on AD
vs MCI, and 99.96% on MCI vs EMCI. The high
accuracy values indicate that the proposed CNN can
effectively differentiate between different stages of
the disease. The CNN proposed model, which
underwent comprehensive training on a dataset
pertaining to Alzheimer’s ailment comprising five
distinct classes. Upon assessment, it demonstrated an
excellent F1 score accuracy of 1.00 as explained in
table 1 and a remarkable ROC AUC score of 0.9999
as shown in Fig. 6. Furthermore, it also displayed
admirable precision values of 0.9989 along with
substantial recall and sensitivity scores of 0.9958.
Finally, its resulting test accuracy amounted to an
impressive 99.92% of the system as shown in Fig. 8.
According to the Fig. 12, Fig. 13, and the confusion
matrix in Fig. 14.
Fig. 12. VGG19 accuracy
Fig. 13. VGG19 loss
It became rather obvious that the VGG19 model is
failing in producing any accurate predictions across
any available classes. To be precise, all samples from
every class have been incorrectly predicted as
belonging to an entirely separate class (class 0).
Ultimately this led to a diagnostic diagram featuring
values placed only along its main diagonal and
corresponding strictly to their respective total sample
count within different classes.
Fig. 14. VGG19 Confusion Matrix
The (PPV) and (NPV) cannot be calculated because
all the predicted values for each class are 0, except
for the true positives on the diagonal as shown in
Fig. 15.
Fig. 15. ROC AUC of VGG19
Based on the evaluation of table 2 classification
report surrounding VGG19’s capabilities regarding
task at hand showcased noticeable deficiencies in
performance quality standards specifically relating
to several shortcomings during precision-recall
assessment tests for all assigned categories
incorporating considerably lower than expected f1
scores concerning prediction effectiveness hence
underscoring these models’ inadequacies
furthermore while providing additional insight into
its low levels of accuracy observed at a paltry rate
of only 21% able to provide accurate forecasts
reflecting insufficient predictive capacity resultant
from inadequate results witnessed on both macro
and weighted average performance evaluation
metrics reflecting the model’s subpar level of
effectiveness across assigned categories with an
overarching macro-average f1-score of just 0.07
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.13
Asmaa Nasr Mohammed,
Abdulgani Albagul, Moamer Musbah Ahmad
E-ISSN: 2224-3488
123
Volume 19, 2023
attesting to this observation.
Table 2: Vgg19 classification report
Precision
Recall
f1-score
Support
LMCI (0)
0.21
1.00
0.34
1662
EMCI (1)
0.00
0.00
0.00
1612
MCI (2)
0.00
0.00
0.00
1589
AD (3)
0.00
0.00
0.00
1542
CN (4)
0.00
0.00
0.00
1611
Accuracy
0.21
8016
Macro avg
0.04
0.20
0.07
8061
Weighted avg
0.04
0.21
0.07
8061
According to Fig. 16 and Fig. 17, the model learns to
recognize patterns in the images, its accuracy im-
proves and its loss decreases.
Fig. 16: Model accuracy
Fig. 17. Model loss
The ResNet50’s test accuracy recorded stood at an
excellent figure at 91.31% is also extremely high,
showing that the model is performing well on the test
set with a decreased loss recorded at around 0.323.
From Fig, 18, The overall PPV of 0.923 means that
out of all the samples that were predicted to be
positive, 92.3% of them were positive and correctly
classified. On the other hand, the overall NPV of 0.98
means that out of all the samples that were predicted
to be negative, 98% of them were negative and
correctly classified. These values show that the
model has a high degree of accuracy in predicting
both positive and negative classes.
Fig. 18. ResNet50 Confusion Matrix
According to Table 3’s classification report, it
appears that the ResNet50 outperformed the VGG19
on our current task at hand by a significant margin.
Across all categories, values for precision, recall, and
f1-score were significantly superior within the
ResNet50 model - suggesting its superior accuracy
when making forecasts. Moreover, with an accuracy
of 91%, it is correct with almost every nine out of ten
predictions made by this model implemented
correctly! Furthermore, measuring against macro-
average and weighted-average metrics across all
categories also indicates progressively better results
for ResNet50 than for its counterpart - VGG19.
Lastly mentioned is a strong indication towards an
excellent overall summation characterizing
performance results by presenting a Macro- Average
F1-Score of 0.91.
Table 3: ResNet50 Classification report
Precision
Recall
f1-score
Support
LMCI (0)
0.87
0.90
0.88
1662
EMCI (1)
0.89
0.89
0.89
1612
MCI (2)
0.90
0.94
0.92
1589
AD (3)
0.93
0.97
0.95
1542
CN (4)
0.98
0.87
0.92
1611
Accuracy
0.91
8016
Macro avg
0.92
0.91
0.91
8061
Weighted avg
0.91
0.91
0.91
8061
According to our findings, while analyzing three
models namely, VGG19, ResNet50 and our newly
introduced CNN architecture, we discovered that
VGG19 had a higher loss value compared to its
counterparts indicating lower accuracy in predictions
(with an overall loss value of 1.6093) as shown in Fig
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.13
Asmaa Nasr Mohammed,
Abdulgani Albagul, Moamer Musbah Ahmad
E-ISSN: 2224-3488
124
Volume 19, 2023
19.
Fig. 19: ResNet50 ROC curve
One possible explanation behind this discrepancy
previously observed by computer vision researchers
ould stem from the depth and complexity featured
within VGG19’s design as it has many parameters
and hence may be prone to over-fit. On another note,
having described fewer complex techniques applied
in reducing data variances for regression problems
i.e., applying regularization techniques such as
embedding dropout layers within modeling
architectures amongst others; The use residual
connections present in ResNet50 may account for
why ResNet50 displayed better forecasting ability
with a decreased loss recorded at around 0.323 as
shown in Fig. 17. The proposed CNN architecture
yielded even better prediction ability with a low-
loss value at about 0.006862 as shown in Fig. 9,
setting an impressive standard for true precision,
hence championing it as a preferred choice for the
task at hand. The exceptional performance of the
CNN model that has been suggested on the
Alzheimer’s dataset with five classes can be traced
back to its specificity to this dataset. As it has been
trained specifically using this dataset, it may have
learned unique features that are specific to this task.
These acquired features have resulted in its
outstanding F1 score accuracy, high ROC AUC
score, recall, sensitivity, precision, specificity, and
test accuracy. The less-than-ideal results generated
by VGG19’s classification attempts on Alzheimer’s
dataset are seemingly rooted in its pre-training on
ImageNet dataset. The dissonance regarding class
and feature factors between these two datasets meant
VGG19 lacked adequate requisite-feature knowledge
needed for precise classifications under current five-
class diagnosis models. While ResNet50 excelled
beyond VGG19, its performance pales when
evaluated against that of the proposed CNN
counterpart precisely tailored for accommodating a
forementioned specifications specific to Alzheimer’s
dataset with five classes. To sum up, the results from
the Alzheimer’s dataset with five classes indicate that
the CNN model surpasses pre-trained models like
VGG19 and ResNet50 in terms of F1 score accuracy,
ROC AUC score, recall, sensitivity, precision,
specificity, and test accuracy.
4 Conclusion
The proposed CNN model demonstrated exceptional
accuracy and sensitivity in detecting AD and exhibited
an impeccable test accuracy score of 99.92%. The
model’s ROC AUC score of 0.9999 indicates that it
can effectively distinguish between positive and
negative cases of AD. PPV and NPV for all classes
are 1.00 which indicate a high accuracy for both
positive and negative predictions. The proposed
model outperformed pre-trained models VGG19 and
ResNet50 in all measures, with a low-loss value and
exceptional F1 score accuracy, ROC AUC score,
recall rate, precision, specificity, and test accuracy. In
contrast, VGG19 showed poor performance with F1
grade accuracy of 0.21, ROC AUC score of 0.5, recall
of 0.2, and low accuracy, resulting in a test accuracy
of only 20.73%. ResNet50 performed better than
VGG19, but the accuracy of the F1 score of 0.91 and
the ROC AUC score of 0.99374 were lower than that
of the proposed CNN model. The proposed CNN
model could recognise the specific brain regions such
as the corpus callosum and thalamus which play a
significant role in identifying Alzheimer’s disease
images and acknowledges their importance as feature
maps. In general, it can be notice that MRI and CNN
models hold promise in improving both the accuracy
and efficiency of AD diagnosis. Further exploration
in this field may result in the development of a
standardized approach to diagnosing AD ultimately
benefiting individuals who suffer from it.
References:
[1] Alzheimer’s Disease and Dementia. What is
Alzheimer’s?[Online].Available:https://www.al
z.org/alzheimers-dementia/what-is-alzheimers.
[Accessed: Dec. 21, 2022].
[2] J. Hardy and D. J. Selkoe,” The amyloid
hypothesis of Alzheimer’s disease: Progress and
problems on the road to therapeutics,” Science,
vol. 297, pp. 353-356, 2002. doi:
10.1126/science.1072994.
[3] J. Cummings, G. Lee, A. Ritter, M. Sabbagh,
and K. Zhong,” Alzheimer’s disease drug
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.13
Asmaa Nasr Mohammed,
Abdulgani Albagul, Moamer Musbah Ahmad
E-ISSN: 2224-3488
125
Volume 19, 2023
development pipeline: 2019,” Alzheimer’s
Dementia: Translational Research Clinical
Interventions, vol. 5, pp. 272-293, 2019. doi:
10.1016/j.trci.2019.05.008.
[4] R. F. Gottesman, A. L. C. Schneider, Y. Zhou,
J. Coresh, E. Green, N. Gupta, D. S. Knopman,
A. Mintz, A. Rahmim, A. R. Sharrett, L. E.
Wagenknecht, D. F. Wong, and T. H. Mosley,”
Association Between Midlife Vascular Risk
Factors and Estimated Brain Amyloid
Deposition,” JAMA, vol. 317, no. 14, pp. 1443-
1450, 2017. doi: 10.1001/jama.2017.3090.
[5] G. Livingston, A. Sommerlad, V. Orgeta, S. G.
Costafreda, J. Huntley, D. Ames, C. Ballard, S.
Banerjee, A. Burns, J. Cohen-Mansfield, C.
Cooper, N. Fox, L. N. Gitlin, R. Howard, H. C.
Kales, E. B. Larson, K. Ritchie, K. Rockwood,
E. L. Sampson, Q. Samus, L. S. Schneider, G.
Selbæk, L. Teri, and N. Mukadam, ”Dementia
prevention, intervention, and care,” The Lancet,
vol. 390, no. 10113, pp. 2673-2734,
2017.doi:10.1016/S0140-6736(17)31363-6.
[6] P. B. Rosenberg, M. M. Mielke, B. S. Appleby,
E. S. Oh, Y. E. Geda, and C. G. Lyketsos,” The
association of neuropsychiatric symptoms in
MCI with incident dementia and Alzheimer
disease,” The American Journal of Geriatric
Psychiatry, vol. 26, no. 10, pp. 1015-1023, 2018.
doi: 10.1016/j.jagp.2018.05.002.
[7] D. W. Shattuck and R. M. Leahy,” Automated
graph-based analysis and correction of cortical
volume topology,” IEEE Transactions on
Medical Imaging, vol. 20, no. 11, pp. 1167-
1177, 2001. doi: 10.1109/42.963819.
[8] S. M. Smith,” Fast robust automated brain
extraction,” Human Brain Mapping, vol. 17, no.
3, pp. 143-155, 2002. doi: 10.1002/hbm.10062.
[9] C. Fennema-Notestine, I. B. Ozyurt, K. A.
Clark, S. Morris, A. Bischoff-Grethe, M. W.
Bondi, T. L. Jernigan, B. Fischl, F. Segonne, D.
W. Shattuck, R. M. Leahy, D. E. Rex, A. W.
Toga, K. H. Zou, and G. G. Brown,
”Quantitative evaluation of automated skull-
stripping methods applied to contemporary and
legacy images: Effects of diagnosis, bias
correction, and slice location,” Human Brain
Mapping, vol. 27, no. 2, pp. 99-113, 2006. doi:
10.1002/hbm.20161.
[10] R. Li, W. Zhang, H. I. Suk, L. Wang, and D.
Shen,” Deep learn- ing based imaging data
completion for improved brain disease di-
agnosis,” Medical Image Analysis, vol. 63, p.
101691, 2020. doi:
10.1016/j.media.2020.101691.
[11] Y. LeCun, Y. Bengio, and G. Hinton,” Deep
learning,” Nature, vol. 521, no. 7553, pp. 436-
444, 2015. doi: 10.1038/nature14539.
[12] I. Goodfellow, Y. Bengio, and A. Courville,”
Deep Learning,” MIT Press, 2016.
[13] A. Krizhevsky, I. Sutskever, and G. E. Hinton,”
ImageNet Classification with Deep
Convolutional Neural Networks,” in Advances
in Neural Information Processing Systems, vol.
25, pp. 1097-1105, 2012.
[14] K. He, X. Zhang, S. Ren, and J. Sun,” Deep
Residual Learning for Image Recognition,” in
Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pp.
770-778, 2016. doi: 10.1109/CVPR.2016.90.
[15] J. Long, E. Shelhamer, and T. Darrell,” Fully
Convolutional Networks for Semantic
Segmentation,” in Proceedings of the IEEE
Conference on Computer Vision and Pattern
Recognition, pp. 3431-3440, 2015. doi:
10.1109/CVPR.2015.7298965.
[16] Y. Kim,” Convolutional neural networks for
sentence classification,” in Proceedings of the
Conference on Empirical Methods in Natural
Language Processing, pp. 1746-1751, 2014. doi:
10.3115/v1/D14-1181.
[17] K. Zhang, Z. Zhang, Z. Li, and Y. Qiao,” Joint
face detection and alignment using multitask
cascaded convolutional networks,” IEEE Signal
Processing Letters, vol. 23, no. 10, pp. 1499-
1503, 2017. doi: 10.1109/LSP.2016.2640551.
[18] Y. Zhang, H. Chen, L. Chen, and T. Huang,”
Joint training of cascaded CNN for face
detection,” Pattern Recognition, vol. 74, pp.
308-316, 2018. doi:
10.1016/j.patcog.2017.10.005.
[19] V. Mnih, K. Kavukcuoglu, D. Silver, A. A.
Rusu, J. Veness, M. G. Bellemare, A. Graves,
M. Riedmiller, A. K. Fidjeland, G. Ostrovski, S.
Petersen, C. Beattie, A. Sadik, I. Antonoglou, H.
King, D. Kumaran, D. Wierstra, S. Legg, and D.
Hassabis,” Human-level control through deep
reinforcement learning,” Nature, vol. 518, no.
7540, pp. 529-533, 2015. doi:
10.1038/nature14236.
[20] D.R. Sarvamangala and R.V. Kulkarni,”
Convolutional neural networks in medical
image understanding: a survey,” Evolutionary
Intelligence, vol. 15, pp. 1-22, 2022. DOI:
10.1007/s12065-020-00540-3.
[21] A. Singh,” Introduction to neural network:
Convolutional Neural Network,” Analytics
Vidhya, 2020. [Online]. Available:
https://www.analyticsvidhya.com/blog/2020/02
/mathematics-behind- convolutional-neural-
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.13
Asmaa Nasr Mohammed,
Abdulgani Albagul, Moamer Musbah Ahmad
E-ISSN: 2224-3488
126
Volume 19, 2023
network/?utmsource = learn − image −
classification − cnn − convolutional − neural −
networks − 5 − datasets. [Accessed: Dec.21,
2022].
[22] J. Brownlee,” Transfer learning in Keras with
computer vision models,” Machine Learning
Mastery, 2020. [Online]. Available:
https://machinelearningmastery.com/how-to-
use-transfer-learning-when-developing-
convolutional-neural-network-models
[Accessed: May 25, 2023].
[23] A. Rojatkar,” Precision, Recall, Sensitivity and
Specificity,” OpenGenus IQ: Computing
Expertise Legacy, Nov. 4, 2021. [Online].
Available: https://iq.opengenus.org/ precision-
recall-sensitivity-specificity/
[24] S. R. Steadman, M. D. Robertson, and M. J.
Morrell,” Performance Evaluation Metrics for
Convolutional Neural Networks: Overview and
Systematic Eval- uation,” IEEE Access, vol. 7,
pp. 122,190-122,205, 2019. doi: 10.1109/AC-
CESS.2019.2931563.
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The authors equally contributed in the present
research, at all stages from the formulation of the
problem to the final findings and solution.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
No funding was received for conducting this study.
Conflict of Interest
The authors have no conflicts of interest to declare
that are relevant to the content of this article.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.13
Asmaa Nasr Mohammed,
Abdulgani Albagul, Moamer Musbah Ahmad
E-ISSN: 2224-3488
127
Volume 19, 2023
