Deep Learning based-stage classification of Alzheimer’ Disease

from MRI
Iheb ELGHAIEB1*, Hiba MZOUGHI 2, Ahmed ZOUINKHI 1 and Naceur ABDELKRIM1, *
1
University of Gabes, National School of Engineers of Gabes.MACS-lab
Systems, St Omar Ibn El Khattab, 6029 Gabes, Tunisia

2
National engineering School of Sfax, ATMS-Lab, – Sfax
II. RELATED WORKS
Abstract— This study proposes a pretrained weight transfer
Convolutional Neural Networks (CNNs), one type of deep
*[email protected]
learning technique using popular pretrained convolutional
learning technology, have shown impressive results in a variety
*[email protected]
neural networks such as Vgg16 and Efficientnet-B3. We
of picture categorization applications, especially in recent years.
also used a classical CNN architecture with domain-specific
With only the raw visual input, CNNs may automatically
information and architectural alterations to evaluate the
generate hierarchical representations, capturing discriminative
performance of the suggested pretrained models. The
features necessary for precise classification.
outcomes of the trials show that the transfer learning
method performs better in classifying Alzheimer's disease
than the suggested CNN, efficientnetb3, and Vgg16 the
potential of transfer learning to enhance the early detection The use of deep learning techniques and transfer learning
and treatment of the condition. The suggested models algorithms has led to a major advancement in the field of
display results with an overall accuracy of 95.70% for CNN, Alzheimer disease (AD) classification using MRI images in
Vgg16 and Efficientnet-b3. recent years. The potential advantages of these techniques for
enhancing AD classification accuracy and early detection have
Keywords— Alzheimer Disease, MRI Images, Alzheimer’s been the subject of numerous investigations. An overview of the
key studies on this subject is given in this review of the literature.
stage classification,
Nawaz et al. [2] Based on deep characteristics, a real-time
I. INTRODUCTION approach for identifying Alzheimer disease stages is proposed.
The scientists use a combination of manually developed and
One prevalent kind of dementia that impairs memory and deep learning-based characteristics extracted from MRI scans
cognition is Alzheimer's disease (AD). Significant public to classify different stages of Alzheimer disease. Their
health issues have been raised by Alzheimer's disease and other approach produces promising results in terms of accuracy and
related dementias in North America and Europe. According to instantaneous operation.
estimates, 10% of people over 65 in North America and Europe
had Alzheimer's disease in 2019. That comes to about 5.6
million people. With the aging of the population and the Bhandari et al. [3] demonstrated how to use segmentation
anticipated sharp growth in the number of Alzheimer's algorithms to analyze brain subregions and predict AD in its early
patients, this figure is predicted to rise. Most of the care stages. The MRI dataset used in the study includes patients with
required by individuals with Alzheimer's disease is given by AD as well as those with normal cognitive function. During the
friends, family, and other informal caregivers [1]. AD is MRI processing, skull stripping and histogram equalization were
commonly divided into four phases: Non Demented Very Mild applied to improve the image quality. The multilevel
Demented Mild Demented Moderate Demented thresholding technique was then applied to the MRI's segment-
specific regions, including the white matter, corpus callosum,
gray matter, and hippocampus sections. The resulting segmented
images were placed into a CNN to predict AD. The proposed
framework improved CNN performance by highlighting AD-
related brain areas on MRI using multilayer thresholding.
Regretfully, the automatic thresholding approach relied on a
strong contrast between the backdrop and the applied
segmentation algorithm, which was not integrated with the CNN.

Altaf et al. [4] provide a hybrid feature extraction technique

for the multi-class categorization of Alzheimer's disease from
MRI images. The authors combine shape-, texture-, and
Figure1: four phases in AD (Non Demented , Very Mild volumetric features in a support vector machine classifier. The
Demented , Mild Demented , Moderate Demented) study demonstrates the efficacy of hybrid features in the
multiclass classification of Alzheimer's disease.
The use of magnetic resonance imaging, or MRIs, as a research
tool to examine the anatomical alterations in the brain linked
to AD has increased. However, given the complexity and N. M. Khan et al [5] utilized brain MRI to apply transfer
considerable variation in brain structures, identifying AD learning. Initially, they trained numerous CNNs in binary and
patients based solely on Manuel's interpretation of MRI data multiple class classification using the same dataset. The
presents many difficulties. Recurrent Neural Network (RNN) was then fed features from the
CNN model. Despite recent successes in computer science Three goals are to be achieved in this work: (1) assess Vgg16's
with deep learning, these algorithms are limited in the number effectiveness in classifying AD on MRI images; (2) suggest new
of training images they can handle. CNN and efficientnetb3 architectures that are better suited for
AD classification; and (3) offer insights into the potential of
Hessam Ahmadi et Al [6] Alzheimer's disease (AD)
transfer learning for early AD diagnosis and intervention.
develops years before dementia and other associated signs and
symptoms show up. Early detection of AD is crucial to slowing
the rate of brain degeneration. In the whole-brain analysis,
III. Methodology
none of the research groups showed any significant alterations
Using pretrained networks, we employed Transfer learning to
over the course of the disease's one-year progression.
speed the training of neural network models. With pretrained
Nonetheless, the intergroup comparison demonstrated that the
networks that are tailored for specific tasks on larger datasets, this
brain network's efficiency, transitivity, modularity, and
method can nevertheless yield valuable results in situations
clustering coefficient had all altered dramatically across the
where training data is scarce. Convolutional Neural Networks
stages (from healthy to AD). The DMN network analysis
(ConvNets) are widely used for medical image processing, as
revealed notable modifications in the EMCI participants, but
seen in Figure 1. VGG 16 and efficientnetb3 were chosen for this
no significant changes in the other study groups. Participation,
study.
clustering, and degree were three metrics in the nodal analysis
By accurately identifying previously unidentified fracture
of the DMN that varied considerably as AD progressed.
images, the trained transfer learning model generates predictions
for fracture classes in real-world scenarios. The detailed phases
of the transfer learning models offer a comprehensive foundation
Xiaocai Shan [7] The statistical dependency of information for Alzheimer's stage classification. Conversely, as the results
flow between cortical regions, or functional connectivity, in section explains, the proposed CNN uses hyper parameter tuning
the human brain makes a substantial contribution to our to perform specific layer alterations.
understanding of the intrinsic brain network and its functioning The suggested CNN, when used in conjunction with
mechanism. In order to optimize classification performance,
efficientnetb3 and Vgg16 transfer learning models, can
this paper presents a novel dynamical spatial–temporal graph
effectively classify Alzheimer's stages. While topologies and
convolutional neural network (ST-GCN) to fully explore its
feature sets differ, the steps for using transfer learning to fracture
potential in the early diagnosis of Alzheimer's disease (AD) classification are the same. Among these are the following:
utilizing electroencephalogram (EEG) recordings. The
proposed ST-GCN takes into account both the corresponding
dynamics of the signal EEG channel and the adjacency matrix
of functional connectivity from multiple EEG channels
simultaneously, which sets it apart from previous studies that
focus on either topological brain function characteristics or
temporal features of EEG.
Daichi Shigemizu et Al [8] Among the elderly, the most
prevalent multifactorial neurodegenerative illness is late-onset
Alzheimer's disease (LOAD). Because LOAD is
heterogeneous, each patient experiences LOAD differently.
Genetic risk factors for LOAD have been found through
genome-wide association studies (GWAS), but not for LOAD
subtypes. Based on Japanese GWAS data from 1947 patients
and 2192 cognitively normal controls in a discovery cohort and
847 patients and 2298 controls in an independent validation
cohort, we investigated the genetic architecture of LOAD in
this study. There were found to be two different categories of
LOAD patients. One was typified by immune-related genes
(RELB and CBLC) and high risk genes for developing LOAD
(APOC1 and APOC1P1). Genes linked to kidney problems
(AXDND1, FBP1, and MIR2278) were found in the other.
Analysis of the albumin and hemoglobin levels obtained from
routine results of a blood test indicated that LOAD
pathogenesis might be caused by compromised kidney Figure 2: Flowchart of the proposed framework for MRI
function.
This work improved the technique for diagnosing AD on MRI
images by presenting a transfer learning mechanism. The A. Data Preprocessing:
proposed research focuses on four extensive architectures: Preprocessing the dataset, which comprises of pictures depicting
CNN, efficientnetb3, Vgg16, and efficientnetb3. various fracture kinds, is necessary before training the transfer
The study looks into how well it works to increase the accuracy learning models. Image scaling, normalization, augmentation
of AD classification and extract useful information from MRI (such as rotation and flipping), and data splitting into training
pictures. Furthermore, we provide domain-specific CNN, and validation sets are examples of preprocessing procedures.
Vgg16, and efficientnetb3 architectures that integrate Image resizing:
architectural alterations with domain-specific information, Image scaling is a crucial pre-processing step in computer
specifically optimized for AD classification. vision. Mostly, deep learning models train faster on small
images. A larger input image requires four times as many pixels
for the neural network to learn from, which extends the
architecture's training time. We showed in this study how
model performance and training time are impacted over time
by image scaling. The 227x227 original size of the input
photographs has been restored to 127x127.
Intensity-Normalization:
The simple technique of min-max normalization uses the
original data's minimum and maximum values to rescale the
data's values to a range between 0 and 1. This technique
preserves the relative order and separation between the data
points while reducing variation and enhancing the effect of
outliers.
Dataset Augmentation:
To expand the amount of a training dataset, fictitious additions
to the existing data are created using methods like rotational
augmentation and flipping. This can enhance machine
learning models' performance and generalization.
Dataset Splitting: Figure 4: Architecture vgg16
Data splitting is the division of data into two or more - EfficientnetB3
subgroups. In a two-part split, the data is usually evaluated or
tested in the first half, and the model is trained in the second The Efficient-Net family of models uses a revolutionary scaling
half. Figure 2 shows the split Alzheimer dataset. technique that uniformly modifies the network's depth, width,
and resolutions. The novel scaling technique delivers higher
performance and lower footprints as compared to ConvNets ,
The architecture of the EfficientNetB3 model is shown in Figure
5. The EfficientNet-B3 concept is based on the Mobile Block
Inverted Bottleneck Conversion. By maintaining the ratio
constant, the layers of this model have been improved, perhaps
improving the model's functionality. The convolutional layer
receives the output of the preceding layer via "average pooling"
and a small, down-sampled (3x3) kernel size. The result is then
received by the Fully Connected (FC) layer, which uses it to
calculate the final classification.

Figure 3: dataset split

B .Pretrained Model Selection:

 VGG|16
The first and second convolutional layers of the VGG-model
have 64 feature kernel filters, each with a 3x3 filter size. If an
RGB image with a depth of three is placed into these layers,
the dimensions become 224x224x64. The output is then sent
to a max pooling layer with a stride of two. In the third and
fourth convolutional layers, filters with a 3x3 filter size and
128 feature kernel filters are employed. After these layers, a
max pooling layer with a stride of 2 is applied, reducing the
output to 56x56x128. Five, Six, and Seven are convolutional
layers with 256 feature mappings and a 3x3 kernel size ,
Following these layers is a max pooling layer with a stride of
Figure 5: Architecture efficientnetb3
two.
The ninth through thirteenth layers are two sets of - CNN
convolutional layers with a 3x3 kernel size. Both sets have an MRI. Slices are obtained using three different picture
512 kernel filters. After convolution, a max pooling layer with planes: the axial or horizontal plane, the coronal or frontal
a stride of 1 comes after these layers. The fourteenth and plane, and the sagittal or center plane. In addition, we added
fifteenth fully connected hidden levels each contain 4096
additional training examples by using Horizontal Flipping.
units. Following them is a softmax output layer (sixteenth
layer) with 1000 units. Neurons in convolutional neural networks have weights
and biases that are tuned to match the different objects in
the image.
The practice of adapting a model trained on one task to
another that is similar is known as transfer learning in
machine learning. By transferring knowledge from one
domain to another, transfer learning can improve the
model's performance on the new task, particularly in
scenarios when training data is scarce.

Figure 7 : Exemple of the dataset from each class

 Evaluation metrics
The validation set is used to assess the transfer learning model's
performance following classifier training. Measures like recall,
accuracy, precision, and F1-score are calculated to evaluate how
well the model performs in categorization. To enhance
performance, the model's parameters can be adjusted further as
necessary:
- Accuarcy (%):
A statistic known as accuracy is used to assess the overall
precision of the model's predictions. It shows the proportion of
the dataset's correctly identified samples. The accuracy of the
model attained after a particular epoch is shown by ACC (%),
which is expressed as a percentage.
𝑇𝑁 + 𝑇𝑃
(1)
𝐹𝑃 − 𝑇𝑃 + 𝐹𝑁 + 𝑇𝑁

- Precision (%):
The model's ability to discriminate positive examples from
expected positives is measured by a statistic called as accuracy.
It represents the ratio of genuine positives to the sum of false
positives and true positives. The percentage P (%) represents
the model's accuracy at the conclusion of a certain period.
Figure 6: Architecture CNN
𝑇𝑃
(2)
𝐹𝑃 + 𝑇𝑃
B. Model Architecture Modification:
The selected pretrained model architecture needs to be
modified for the specific purpose of Alzheimer's disease stage - Recall (%):
classification. The last few layers of the model, including the Recall, also known as true positive rate or sensitivity, assesses
fully connected layers, are typically adjusted in this step to the model's ability to distinguish between positive and true
better match the desired output classes and dimensions. positives. It is calculated by dividing the number of true
positives by the total number of false negatives and positives.
A percentage known as R (%) represents the recall that the
IV. RESULT AND ANALYSIS model achieved at the end of a given epoch.

 Dataset 𝑇𝑃
(3)
The dataset includes preprocessed MRI (Magnetic Resonance 𝑇𝑃𝐹𝑁
Imaging) images. The data was collected from a variety of
websites, healthcare facilities, and free sources. Each image VALIDATION OF THE PROPOSED FRAMEWORK
has been reduced to 128 by 128 pixels. The dataset consists of
six 400 MRI scans divided into four picture groups. a) Accuracy Plot
 Class1: Non Demented (896 pictures) Accustomed to CNN. A three-layer convolutional model is used
 Class 2: Moderate Demented (64 pictures) to convert the input image size (227x227) to (127x127), yielding
 Class 3: Mild Demented (3200 pictures) a feature matrix with 58 dense layers. Figure 8 shows a
 Class 4: Very Mild Demented (2240 pictures) pretrained CNN. The plot's accuracy after 10 epochs is 88.89%.
The loss plot in Figure 8s depicts an exponential decrease in loss
after 8 epochs.
 a) Confusion Matrix
The confusion matrix of the suited models are given in
the figure bellow.

Figure 8: CNN Accuracy Plot

The VGG16 Model generated an output feature matrix
with four dense layers after applying sixteen layers of (A) ( B)
convolution to an input picture with dimensions of 227
× 227. After 25 epochs, the accuracy plot in figure 9 C
shows 85.07%. Figure 9 demonstrates how the loss
diminishes exponentially over ten epochs.

(c)
Figure 11: A: VGG-16, B: CNN, and C: Efficient net’
Confusion Matrix

Figure 9: VGG16 Accuracy plot  Evaluation Results for model :

In this study, precision is defined as the percentage of
Adapted efficientnetb3. The 227x227 input image is demented individuals that were properly predicted,Table
fed into an eight-layer convolution model, and the I demonstrates that the precision values are practically
result is a four-dimensional feature matrix named 745. identical for both architectures. The memory score is the
Here, modified means has reduced the number of percentage of cases of dementia that were predicted
convolutions from five to two. The Adam optimization correctly. There was no statistically significant
changes the optimization function as well. It reaches difference between the recall values of the two
95.70% accuracy after 20 epochs, as shown in Figure architectures, which were nearly identical.
10 of the Accuracy Plot.
Because of the dataset's unequal class distribution,
the accuracy value is not always correct. As a result, an
F1-score value is required to evaluate each architecture's
performance in comparison to its class. This is because
the F1-score is determined by analyzing precision and
recall statistics. We achieved an extremely high F1-
score, virtually perfect, for both designs. The
extraordinarily high accuracy and F1-score demonstrate
the architecture's strong validity.

Figure 10 : efficientnetb3 Accuracy plot

 Table 1 : accuracy measurement. [4] [4] T. Altaf, S. M. Anwar, N. Gul, M. Majid, and M. N. Majeed, “2017
Multi-class Alzheimer Disease Classification using Hybrid Features.pdf,”
Class Precisio recall F1-score Acc 2017 Future Technologies Conference (Ftc), no. November, pp. 264–267,
n urac 2017.
y%
[5] [5] N. M. Khan, N. Abraham, and M. Hon, “Transfer Learning with
Mild 1.00 0.98 0.99 Intelligent Training Data Selection for Prediction of Alzheimer’sDisease,”
Demented IEEE Access, vol. 7, pp. 72726–72735, 2019,
doi:10.1109/ACCESS.2019.2920448.
Moderat 1.00 1.00 1.00
Effici 95.70%
e Demented [6] [6] Hessam ahmadi ,emad fatemizadeh “Identifying brain functional
entne
tb3 0.99
connectivity alterations during different stages of Alzheimer’s disease
Non 0.99 0.99
“Received 06 Apr 2020, Accepted 12 Nov 2020, Published online: 21 Dec
Demented
2022.
Very 0.98 0.99 0.99 [7] [7] Xiaocai Shan, Jun Cao, Shoudong Huo, Liangyu Chen, Ptolemaios
Mild Georgios Sarrigiannis, Yifan Zhao “Spatial–temporal graph convolutional
Demented network for Alzheimer classification based on brain functional
connectivity imaging of electroencephalogram” First published: 25 June
Mild 1.00 0.99 0.99 2022,https://doi.org/10.1002/hbm.25994.
Demented
[8]
85.7
Moderat 1.00 1.00 1.00 [9]
% [8]Daichi Shigemizu, Shintaro Akiyama, Mutsumi Suganuma, Motoki
e Demented Furutani, Akiko Yamakawa, Yukiko Nakano, Kouichi Ozaki , Shumpei
Vgg- Niida” Classification and deep-learning–based prediction of Alzheimer
16 Non 0.99 1.00 0.99 disease subtypes by using genomic data” Published: 29 June 2023
Demented
Very 1.00 0.99 0.99
Mild
Demented
Mild 1.00 0.99 0.99
Demented
CNN
Moderat 1.00 1.00 1.00 88.9
e Demented %
Non 0.99 1.00 0.99
Demented

Very 1.00 0.99 0.99

Mild
Demented

Conclusion

To summarize, the use of transfer learning models has resulted

in considerable advances in Alzheimer stage classification.
Several researches have investigated the use of models like
Vgg16, CNN, and proposed Efficientnetb3 architectures to
increase the accuracy and efficiency of Alzheimer's disease
stage classification. It is critical to remember that the success
of transfer learning models is contingent on the availability of
high-quality, properly annotated fracture datasets. More study
is needed to evaluate the models' efficacy on bigger and more
representative datasets, to enlarge the dataset, and to include
a variety of fracture forms.

REFERENCES

[1] [1] Elghaieb, I., Souid, A., Zouinkhi, A., & Sakli, H. (2024). Defeating
Alzheimer's: AI Perspective from Diagnostics to Prognostics: Literature
Summary. In Machine Learning and Deep Learning Techniques for
Medical Image Recognition (pp. 245-256). CRC Press.
[2] [2] H. Nawaz, M. Maqsood, S. Afzal, F. Aadil, I. Mehmood, and S.
Rho,“A deep feature-based real-time system for Alzheimer disease
stage detection,” Multimedia Tools and Applications, vol. 80, no. 28–
29, pp.35789–35807, 2021, doi: 10.1007/s11042-020-09087-y.
[3] [3] A. K. Bhandari, V. K. Singh, A. Kumar, and G. K. Singh, “Cuckoo
search algorithm and wind driven optimization-based study of satellite
image segmentation for multi-level thresholding using Kapur’s
entropy,” Expert Systems with Applications, vol. 41,no. 7, pp. 3538–
3560, Jun. 2014, doi: 10.1016/j.eswa.2013.10.059.