Improved Alzheimer’s Detection from Brain MRI via Transfer Learning

on Pre-Trained Convolutional Deep Models

Malek Jallali

1 a

, Raouia Mokni

2,3 b

and Boudour Ammar

1,4 c

Research Groups in Intelligent Machines(REGIM), National Engineering School of Sfax, University of Sfax, Tunisia

Advanced Technologies for Environment and Smart Cities(ATES Unit), University of sfax, Tunisia

The Higher Institute of Management of Gab

es(ISGG), University of Gabes, Tunisia

Department of Computer Management at the Higher Business School of Sfax (ESC Sfax), University of Sfax, Tunisia

Keywords:

Alzheimer’s Disease (AD), Convolutional Neural Network (CNN), Deep Learning (DL), MCNN, Transfer

Learning, Fine-Tuned, FT-VGGNet19, Brain MRI Images.

Abstract:

Alzheimer’s Disease (AD) presents a major challenge in modern healthcare due to its complex diagnosis and

management. Early and accurate detection is essential for improving patient care and enabling timely thera-

peutic interventions. Research suggests that neurodegenerative changes associated with AD may appear years

before clinical symptoms, highlighting the need for advanced diagnostic techniques. This study explores deep

learning models for classifying AD stages using MRI scans. Speciﬁcally, we propose a Modiﬁed Convolu-

tional Neural Network (MCNN) and a ﬁne-tuned VGGNet19 (FT-VGGNet19) architecture. Both models were

evaluated on the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset, leveraging data augmentation

to enhance generalization and mitigate dataset limitations. Experimental results show that data augmentation

signiﬁcantly improves classiﬁcation performance. The FT-VGGNet19 model achieved the highest accuracy,

reaching 90% on the original dataset and 92% with augmented data. This study highlights the strengths of

each model for clinical applications, emphasizing the role of optimized deep-learning frameworks in early AD

detection. The ﬁndings contribute to developing robust and scalable diagnostic systems, offering promising

advancements in neurodegenerative disease management.

1 INTRODUCTION

Alzheimer’s disease, the most common form of de-

mentia, is a progressive neurodegenerative disorder

that typically begins with subtle memory loss and ad-

vances to profound cognitive decline, ultimately im-

pairing a person’s ability to communicate coherently

or interact with their environment. The disease pri-

marily affects brain region responsible for memory,

language, and reasoning—such as the hippocampus

and cerebral cortex—disrupting neural pathways and

compromising an individual’s ability to perform rou-

tine daily tasks. Current estimates indicate that 1 in 9

adults aged 65 and older lives with Alzheimer’s, ac-

counting for approximately 11.4% of individuals in

this age group worldwide. Characterized by the accu-

mulation of beta-amyloid plaques and neuroﬁbrillary

https://orcid.org/0009-0008-8167-765X

https://orcid.org/0000-0002-6652-5251

https://orcid.org/0000-0002-4934-1573

tau tangles, Alzheimer’s remains incurable, though

ongoing research seeks to unravel its mechanisms and

develop therapies to slow its devastating progression.

(Balasundaram et al., 2023). The primary cause of

Alzheimer’s disease is the abnormal buildup of pro-

teins in the brain, including beta-amyloid plaques and

tau tangles, which contribute to brain cell death and

the shrinkage of brain tissue. While ongoing research

continues to enhance our understanding, there is cur-

rently no cure for Alzheimer’s. However, medications

and lifestyle interventions can help manage symp-

toms and slow the disease’s progression. Medical

imaging, particularly Magnetic Resonance Imaging

(MRI), plays a crucial role in diagnosing Alzheimer’s

by providing detailed insights into brain structures

and identifying characteristic patterns associated with

the condition (Honig and Chin, 2001). Early detec-

tion and accurate classiﬁcation of Alzheimer’s dis-

ease (AD) are crucial for improving therapeutic out-

comes and effectively managing cognitive decline.

The challenge lies not only in identifying the con-

Jallali, M., Mokni, R., Ammar and B.

Improved Alzheimer’s Detection from Brain MRI via Transfer Learning on Pre-Trained Convolutional Deep Models.

DOI: 10.5220/0013523000003967

In Proceedings of the 14th International Conference on Data Science, Technology and Applications (DATA 2025), pages 431-438

ISBN: 978-989-758-758-0; ISSN: 2184-285X

431

dition in its early stages—when symptoms are mild

and neurological changes are subtle—but also in sys-

tematically categorizing it into distinct stages: Non-

Demented, Very Mild Demented, Mild Demented,

and Moderate Demented. Precise staging allows clin-

icians to implement targeted interventions tailored to

the severity of the disorder, signiﬁcantly enhancing

patient care and outcomes(Rasmussen and Langer-

man, 2019). In this context, advanced deep learn-

ing methods, particularly those leveraging brain MRI

analysis, have proven to be effective tools for de-

tecting early signs of Alzheimer’s disease (AD) and

monitoring its progression. MRI is widely used due

to its unparalleled ability to capture detailed struc-

tural features, such as brain atrophy, a hallmark of

AD pathology (El-Latif et al., 2023a) (Mokni and

Haoues, 2022). The main purpose of this paper is

to investigate the application of deep learning mod-

els for the early detection of Alzheimer’s disease us-

ing MRI scans, classifying it into four stages: Non-

Demented, Very Mild Demented, Mild Demented,

and Moderate Demented. The key contributions of

this work are twofold: (1) the development of a modi-

ﬁed CNN architecture (MCNN) tailored speciﬁcally

for Alzheimer’s diagnosis, and (2) the ﬁne-tuning

of a pre-trained VGGNet19 model through transfer

learning to enhance classiﬁcation accuracy and per-

formance. These approaches independently highlight

the effectiveness of CNN architectural modiﬁcations

and transfer learning techniques in improving the de-

tection and classiﬁcation of Alzheimer’s disease. The

remainder of this paper is structured as follows: Sec-

tion 2 reviews related work, while Section 3 describes

the proposed methodology. Section 4 presents the ex-

perimental results, followed by a discussion and com-

parative evaluation with previous studies in Section

5. Finally, Section 6 concludes the paper and outlines

future research directions.

2 RELATED WORK

In this section, we summarize and review ﬁndings

from research studies on Alzheimer’s disease detec-

tion using deep learning models. Ibrahem M.M.

et al., (Madhat et al., 2024)developed an improved

method for detecting Alzheimer’s disease by combin-

ing DenseNet, VGG16, and ensemble learning. They

evaluated performance using accuracy, precision, re-

call, F1-score, and AUC-ROC on preprocessed MRI

scans divided into four disease stages. Techniques

like transfer learning, ﬁne-tuning, and dropout regu-

larization were used. Using the Alzheimer’s Disease

Neuroimaging Initiative (ADNI) Dataset, the method

achieved The ensemble model reached 94% accu-

racy, outperforming standalone CNN (90%), VGG19

(89%), and DenseNet (86%), with better precision

and recall across all stages, highlighting its promise

for early and reliable Alzheimer’s diagnosis in clin-

ical practice. This highlights the potential of com-

bining advanced architectures with ensemble strate-

gies to improve early and reliable detection in clin-

ical settings. Araqi & Abbas (Araqi and Abbas,

2022) developed a CNN-based deep learning ap-

proach for Alzheimer’s detection using brain MRI,

achieving 90.83% accuracy. Their ﬁndings empha-

sized the importance of image preprocessing in sig-

niﬁcantly improving model performance, suggesting

that combining data reﬁnement with deep learning

enhances early and reliable detection. Ajagbe et

al.,(Ajagbe et al., 2021) designed a framework for

classifying Alzheimer’s disease (AD) stages via MRI.

Their model achieved 94.5% accuracy in distinguish-

ing AD from normal controls, surpassing traditional

diagnostic methods. This underscores the potential

of DCNNs for improving early AD detection in clin-

ical settings. Fathi et al., (Fathi et al., 2024) Fathi et

al.,(Fathi et al., 2024) introduced an ensemble CNN

method for early Alzheimer’s detection using MRI

scans. Tested on the ADNI dataset, their approach

reached 91.2% accuracy, demonstrating its reliability

for diagnosis. Table 1 provides an overview of past

research efforts focused on detecting Alzheimer’s dis-

ease.

Table 1: Summary of the related works focused on

Alzheimer’s disease classiﬁcation.

Proposal Used

Dataset

Method Results

Ibrahem

M.M. et

al., (Mad-

hat et al.,

2024)

ADNI

dataset

CNN

VGG16

DenseNet

Ensemble

learning

90%

89%

86%

94%

Araqi

& Abbas

(Araqi and

Abbas,

2022)

alzheimers-

dataset-4-

class-of-

images

CNN 90.83%

Ajagbe et

al.,(Ajagbe

et al.,

2021)

alzheimers-

dataset-4-

class-of-

images

VGG-19 77.66%

Fathi et

al., (Fathi

et al.,

2024)

ADNI

Dataset

Ensemble

learning

CNN

91.2%

DATA 2025 - 14th International Conference on Data Science, Technology and Applications

432

3 PROPOSED METHODOLOGY

As highlighted in Section 2, various studies have

explored Alzheimer’s Disease (AD) diagnosis us-

ing deep learning. Some approaches, such as those

combining DenseNet, VGG16, and ensemble learn-

ing, have demonstrated strong performance but in-

volve complex architectures. Others have employed

CNNs, Deep Convolutional Neural Networks (DC-

NNs), or ensemble learning of multiple CNNs. In

contrast, our approach focuses on a modiﬁed CNN

architecture (MCNN) and a ﬁne-tuned VGGNet19

(FT-VGGNet19), aiming to enhance classiﬁcation ac-

curacy through architectural modiﬁcations and ﬁne-

tuning rather than relying on extensive ensemble

models. Our approach for classifying Alzheimer’s

disease (AD), as shown in Figure 1, involves ﬁve key

steps: (1) Dataset collection (2) Data augmentation,

(3) Data preprocessing, (4) Deep learning-based pro-

posed models, and (5) Decision Making to classify

AD into four stages.

Figure 1: The overview of the proposed Model.

3.1 Dataset Description

We used a publicly available MRI dataset for

Alzheimer’s disease (AD) (Dubey, 2024), sourced

from Kaggle. This dataset includes images from the

Alzheimer’s Disease Neuroimaging Initiative (ADNI)

and contains 6,400 MRI scans split into four cate-

gories: mild dementia (896 images), moderate de-

mentia (64 images), non-dementia (3,200 images),

and very mild dementia (2,240 images). Examples

from the dataset are shown in Figure 2 (El-Latif et al.,

2023b).

Figure 2: Sample MRI images.

3.2 Dataset Augmentation

Data augmentation is a key technique to improve how

well deep learning models classify Alzheimer’s dis-

ease (AD) using MRI scans. AD datasets are often

small due to limited patient availability, which can

cause models to overﬁt and perform poorly on new

data. Data augmentation tackles this by artiﬁcially ex-

panding the training dataset’s size and variety, helping

models learn stronger features and generalize better

to unseen cases (El-Assy et al., 2024). The original

dataset, from Kaggle’s open-access platform (Uran-

injo, 2024), contains 33,984 MRI scans grouped into

four categories: mild dementia (8,960 images), mod-

erate dementia (6,464 images), non-dementia (9,600

images), and very mild dementia (8,960 images). To

expand the dataset, we applied image transformations

(ﬂipping, rotating...) while keeping key features rele-

vant to Alzheimer’s disease intact. The augmentation

techniques employed include:

-Rotation: Random rotations within a range of [-20°,

20°] to simulate different head orientations.

-Scaling: Size adjustments were applied to account

for MRI dimensional variability.

-Flipping: Horizontal ﬂips were applied to increase

dataset variability while preserving the diagnostic fea-

tures.

-Brightness Adjustment: Variations in brightness

were incorporated to simulate the impact of different

lighting conditions during the image capture process.

-Zoom Transformation: Images were randomly

zoomed within a range of ±20% to simulate variations

in scale. Using these augmentation techniques, we

expanded the dataset and boosted the performance of

our deep learning models in classifying Alzheimer’s

disease (AD). Figure 3 shows the distribution of

Alzheimer’s disease categories within the Mild De-

mented, Moderate Demented, Non-Demented, and

Very Mild Demented.

Improved Alzheimer’s Detection from Brain MRI via Transfer Learning on Pre-Trained Convolutional Deep Models

433

Figure 3: The distribution of Alzheimer’s diseases after data

augmentation.

3.3 Dataset Preprocessing

This section describes the essential preprocessing

steps taken to prepare the data for model training, en-

suring efﬁcient learning on a potentially imbalanced

dataset. First, all images were turned to RGB color.

Next, we bundled them into batches of 32, reshaped

them to 224x224 pixels, and scaled their pixel val-

ues. For better model performance and generaliza-

tion, the dataset was split into three subsets: 70% of

the data was used for training, 20% for testing, and

10% for validation. This three-way split ensures that

the model is evaluated on unseen data, helping to mit-

igate the risk of overﬁtting and ensuring that it per-

forms well on new, unseen data.

3.4 Application of Deep Learning-Based

Proposed Models

This study explores the application of deep learning-

based models for Alzheimer’s disease classiﬁcation,

proposing a novel modiﬁed Convolutional Neural

Network (MCNN) and leveraging transfer learning

with pre-trained models. Speciﬁcally, we ﬁne-tuned

VGGNet19 to develop the FT-VGGNet19 model, op-

timizing its performance for AD classiﬁcation.

3.4.1 Modiﬁed-CNN

A Convolutional Neural Network (CNN) is a deep

learning model speciﬁcally designed to process struc-

tured grid-like data, such as images. Its architec-

ture typically consists of multiple layers, each serv-

ing a distinct purpose: convolutional layers extract

meaningful features, pooling layers reduce dimen-

sionality and computational complexity, fully con-

nected layers make high-level decisions, and acti-

vation functions introduce non-linearity to enhance

learning capabilities. Together, these components en-

able CNNs to efﬁciently capture patterns and im-

prove model performance (Awarayi et al., 2024) (Fki

et al., 2024). In this study, we propose a novel archi-

tecture called Modiﬁed-CNN (MCNN), which intro-

duces several enhancements to the standard CNN de-

sign. The MCNN consists of seven blocks, with the

ﬁrst six containing four layers each: three Conv2D

layers followed by a MaxPooling2D layer. The ﬁnal

block includes an additional Conv2D layer to improve

feature extraction. After these blocks, the model is

compressed, followed by two fully connected (dense)

layers with 512 units each and ReLU activation. A

dropout layer with a rate of 0.3 is applied before the

ﬁnal dense layer, which consists of four units and a

softmax activation function for classiﬁcation. Figure

4 provides a detailed overview of the MCNN archi-

tecture.

Figure 4: Proposed MCNN Model.

3.4.2 Fine-Tuned VGGNet 19 Model

In this paper, we investigate the use of transfer

learning and ﬁne-tuning the pre-trained VGGNet 19

model, called Fine-Tuned VGGNet 19 model (FT-

VGGNet 19). Initially, we leverage the VGG19

model, pre-trained on the ImageNet dataset to obtain

DATA 2025 - 14th International Conference on Data Science, Technology and Applications

434

the trained weights parameters that are used to ex-

tract the pertinent features. It consists of 19 layers,

including convolutional, pooling, and fully connected

layers. The network employs 3×3 convolution ﬁlters

with small strides to extract features from input im-

ages while pooling layers reduce the spatial dimen-

sions of feature maps. The fully connected layers per-

form object classiﬁcation, with a softmax output layer

handling multi-class classiﬁcation. In this study, the

model’s pre-trained weights remain frozen through-

out the training process to maintain the integrity of

these learned features. After that, we created a trans-

fer learning method by modifying the architecture by

ﬁrstly ﬂattening, followed by adding a dense layer

and introducing a softmax activation function to clas-

sify the images into four distinct categories related to

Alzheimer’s disease. This adjustment effectively tai-

lors the model for Alzheimer’s disease image classi-

ﬁcation. Figure 5 provides a detailed overview of the

FT-VGGNet19 architecture.

Figure 5: Proposed Fine-Tuned VGGNet19 Model.

4 EXPERIMENTAL RESULTS

In our study, the processing and classiﬁcation tasks

were performed using the Python programming lan-

guage. The models were trained on Google Colab,

leveraging a graphics processing unit (GPU) to en-

hance computational efﬁciency. To investigate the

contribution of deep learning models in classifying

Alzheimer’s disease, we conducted several experi-

ments using the proposed models: Modiﬁed CNN

(MCNN) and Fine-Tuned VGGNet19. Table 2 sum-

marizes the evaluation results for Alzheimer’s dis-

ease classiﬁcation, highlighting the impact of dif-

ferent original data methods and augmentation tech-

niques on accuracy and robustness. The results pre-

sented in Table2 highlight the impact of data augmen-

tation on the performance of deep learning models

for Alzheimer’s disease classiﬁcation. Both the Fine-

Tuned VGGNet19 (FT-VGGNet19) and the Modiﬁed

CNN (MCNN) models demonstrate improved accu-

racy when trained with data augmentation. Specif-

ically, FT-VGGNet19 achieves an accuracy of 92%

with data augmentation compared to 90% without,

while MCNN shows a signiﬁcant improvement from

73% to 90%. Similarly, the F1-score for MCNN im-

proves signiﬁcantly from 64% to 90% with data aug-

mentation, demonstrating a substantial enhancement

in the model’s overall classiﬁcation performance, par-

ticularly in balancing precision and recall. These ﬁnd-

ings emphasize the effectiveness of data augmenta-

tion in boosting model robustness and classiﬁcation

performance, particularly for the MCNN model. In

this study, we conducted two experiments: the ﬁrst

focused on classifying Alzheimer’s disease using the

original data, while the second involved applying data

augmentation to improve classiﬁcation performance

from MRI images.

4.1 The First Experiment

In this section, a deep learning model was used

to classify MRI images into four categories: Mild

Demented, Moderately Demented, Non-Demented,

and Very Mild Demented, using the original dataset.

Model performance was evaluated with a confusion

matrix, comparing predicted labels to actual ones

(Vengala, 2024) and the training and validation ac-

curacy and curves. Figure 6 presents the results of the

confusion matrix for the classiﬁcation of Alzheimer’s

diseases according to four categories obtained from

MCNN and FT-VGGNet19 using the original dataset.

As illustrated in this Figure, FT-VGGNet19 achieved

the highest accuracy, outperforming MCNN in terms

of classiﬁcation effectiveness. Diagonal elements

indicate correct predictions, while off-diagonal val-

ues represent misclassiﬁcations. FT-VGGNet19 cor-

rectly classiﬁes 161 Non Demented and 122 Very

Mild Demented cases, with misclassiﬁcations mainly

occurring between similar classes. Figure 7 pres-

nts the Training /Validation Accuracy curves results

for the classiﬁcation of Alzheimer’s diseases accord-

ing to four categories obtained from MCNN and FT-

VGGNet19 over 15 epochs using the original dataset.

Figure 7 (B) tracks training and validation accuracy

obtained from FT-VGGNet19 across 15 epochs. A

well-trained model shows both curves rising and sta-

bilizing. A divergence where training accuracy in-

creases but validation accuracy stagnates suggests

overﬁtting, while low accuracy in both indicates un-

derﬁtting. The validation curve is crucial for assess-

ing real-world performance.

Improved Alzheimer’s Detection from Brain MRI via Transfer Learning on Pre-Trained Convolutional Deep Models

435

Table 2: Evaluation of Alzheimer’s Disease Classiﬁcation Using the Proposed MCNN Model and Fine-Tuned VGGNet19

Models with Data Augmentation Techniques and Original Data.

Model Data Augmentation Original Data

FT-VGG-Net19 MCNN FT-VGG-Net19 MCNN

Accuracy 92% 90% 90% 73%

Precision 91% 91% 92% 83%

Recall 92% 90% 94% 58%

F1 Score 91% 90% 93% 64%

Figure 6: Confusion Matrices obtained by (A) MCNN (B)

FT-VGGNet19 with original data.

Figure 7: Training and Validation Accuracy curves obtained

by (A) MCNN (B) FT-VGGNet19 with original data.

4.2 The Second Experiment

In this section, the proposed MCNN and Pretrained

FT-VGGNet19 was implemented to classify MRI

scans into four categories: Mild Demented, Moder-

ately Demented, Non-Demented, and Very Mild De-

mented using the augmented dataset to enhance the

model’s robustness and generalization. The model’s

effectiveness was assessed through the confusion ma-

trix, which compares predicted labels with actual

ground truth labels and the training and validation ac-

curacy curves.

Figure 8: Confusion Matrices obtained by (A) MCNN (B)

FT-VGGNet19 with data augmentation.

Figure 8 shows the results of the confusion ma-

trix for the classiﬁcation of Alzheimer’s disease ac-

cording to four categories obtained from MCNN

and FT-VGGNet using the augmented dataset. As

shown in this ﬁgure, using the augmented dataset,

FT-VGGNet19 also achieved the highest accuracy

compared to MCNN. FT-VGGNet19 correctly clas-

siﬁes 508 Mild Demented, 376 Moderate Demented,

DATA 2025 - 14th International Conference on Data Science, Technology and Applications

436

Figure 9: The Training and Validation Accuracy curves ob-

tained by (A) MCNN (B) FT-VGGNet19 with data augmen-

tation.

490 Non-Demented, and 510 Very Mild Demented

cases, with misclassiﬁcations mainly occurring be-

tween similar classes.

Figure 9 illustrates the training and validation ac-

curacy curves for Alzheimer’s disease classiﬁcation

into four categories using MCNN and FT-VGGNet19

over 15 epochs with the augmented dataset. Speciﬁ-

cally, Figure 9 (B) depicts the accuracy trends for the

proposed FT-VGGNet19 model.

A well-trained model exhibits steadily increasing

and stabilizing curves, indicating effective learning.

However, if the validation accuracy plateaus with no-

ticeable ﬂuctuations while the training accuracy con-

tinues to rise, it suggests overﬁtting.

5 DISCUSSION AND

COMPARATIVE EVALUATION

This section presents the results of classifying

Alzheimer’s disease into four categories—non-

demented, very mildly demented, mildly demented,

and moderately demented—using various architec-

tures. We modiﬁed a basic CNN and ﬁne-tuned a pre-

trained VGGNet19, evaluating them on both original

and augmented datasets. As shown in Table 2, data

augmentation had a signiﬁcant positive impact.

The FT-VGGNet19 maintained high performance

Table 3: Comparative analysis of proposed work with pre-

vious works.

Authors Datasets Models Accuracy

(%)

Proposed

alzheimers-

dataset-4-

class-of-

images

FT-

VGGNet19

90%

Augmented

Alzheimer

MRI

Dataset

FT-

VGGNet19

92%

Ajagbe.

et al.,

(Ajagbe

et al.,

2021)

alzheimers-

dataset-4-

class-of-

images

VGG-19 77.66%

Ibrahem

M.M. et

al., (Mad-

hat et al.,

2024)

MRI

dataset

CNN

VGG16

DenseNet

Ensemble

90%

89%

86%

94%

Araqi

& Abbas

(Araqi and

Abbas,

2022)

alzheimers-

dataset-4-

class-of-

images

CNN 90.83%

across both datasets, while the modiﬁed CNN ben-

eﬁted most from augmentation. These results high-

light the value of data augmentation, especially for

custom architectures when data is limited. To the best

of our knowledge, no prior work has been conducted

using augmented data for Alzheimer’s disease classi-

ﬁcation with this dataset. Therefore, we compare our

results with previous studies that used the same orig-

inal dataset. Table 3 highlights the effectiveness of

our proposed model in classifying Alzheimer’s dis-

ease compared to previous studies, including those

by Ajagbe et al. (Ajagbe et al., 2021) and Araqi &

Abbas (Araqi and Abbas, 2022), which were con-

ducted on the original Alzheimer’s dataset (4-class

classiﬁcation). As illustrated in Table 3, our model

achieved an accuracy of 90%, outperforming Ajagbe

et al. (Ajagbe et al., 2021) (77.66%) with an improve-

ment of 12.34%, and yielding comparable results to

Araqi & Abbas (Araqi and Abbas, 2022) (90.83%).

6 CONCLUSION AND FUTURE

WORK

Alzheimer’s disease (AD) is a major neurodegenera-

tive disorder, where early diagnosis is crucial for ef-

Improved Alzheimer’s Detection from Brain MRI via Transfer Learning on Pre-Trained Convolutional Deep Models

437

fective intervention. Traditional diagnostic methods

rely on clinical expertise, which can lead to delays

and inconsistencies. Advances in deep learning, par-

ticularly in medical imaging, have signiﬁcantly im-

proved diagnostic accuracy using brain MRI scans.

This study proposes a modiﬁed CNN (MCNN) and

a ﬁne-tuned VGGNet19 (FT-VGGNet19) for classi-

fying Alzheimer’s disease into four categories—non-

demented, very mildly demented, mildly demented,

and moderately demented, evaluating both on original

and augmented datasets. The FT-VGGNet19 consis-

tently maintained high performance, achieving 92%

accuracy with data augmentation and 90% without. In

contrast, the MCNN beneﬁted signiﬁcantly from aug-

mentation, demonstrating notable improvement. Ad-

ditionally, we assessed the models using precision, re-

call, and F1-score, further validating their effective-

ness. Overall, this study underscores the potential

of deep learning in improving AD diagnosis through

MRI analysis. Future work will explore advanced

augmentation techniques and explainable AI frame-

works to enhance model interpretability and clinical

applicability.

ACKNOWLEDGEMENTS

The research leading to these results has received

funding from the Ministry of Higher Education and

Scientiﬁc Research of Tunisia under grant agreement

number LR11ES48.

REFERENCES

Ajagbe, S. A., Amuda, K. A., Oladipupo, M. A., Oluwaseyi,

F. A., and Okesola, K. I. (2021). Multi-classiﬁcation

of alzheimer disease on magnetic resonance images

(mri) using deep convolutional neural network (dcnn)

approaches. International Journal of Advanced Com-

puter Research, 11(53):51.

Araqi, Z. S. and Abbas, H. H. (2022). Alzheimer’s disease

detection using deep learning on mri images. arXiv,

abs/2204.00068.

Awarayi, N. S., Twum, F., Hayfron-Acquah, J. B., and

Owusu-Agyemang, K. (2024). A bilateral ﬁltering-

based image enhancement for alzheimer disease clas-

siﬁcation using cnn. Plos one, 19(4):e0302358.

Balasundaram, A., Srinivasan, S., Prasad, A., Malik, J., and

Kumar, A. (2023). Hippocampus segmentation-based

alzheimer’s disease diagnosis and classiﬁcation of mri

images. Arabian Journal for Science and Engineer-

ing, 48(8):10249–10265.

Dubey, S. (2024). Search results for alzheimer’s disease

datasets. https://www.kaggle.com/datasets/tourist55/

alzheimers-dataset-4-class-of-images. Accessed:

2024-05-22.

El-Assy, A. M., Amer, H. M., Ibrahim, H. M., and Mo-

hamed, M. A. (2024). A novel cnn architecture for ac-

curate early detection and classiﬁcation of alzheimer’s

disease using mri data. Scientiﬁc Reports, 14(1):3463.

El-Latif, A. A. A., Chelloug, S. A., Alabdulhaﬁth, M.,

and Hammad, M. (2023a). Accurate detection of

alzheimer’s disease using lightweight deep learning

model on mri data. Diagnostics, 13(7):1216.

El-Latif, A. A. A., Chelloug, S. A., Alabdulhaﬁth, M.,

and Hammad, M. (2023b). Accurate detection of

alzheimer’s disease using lightweight deep learning

model on mri data. Diagnostics, 13(7).

Fathi, S., Ahmadi, A., Dehnad, A., Almasi-Dooghaee,

M., Sadegh, M., and for the Alzheimer’s Dis-

ease Neuroimaging Initiative (2024). A deep

learning-based ensemble method for early diagnosis

of alzheimer’s disease using mri images. Neuroinfor-

matics, 22(1):89–105.

Fki, Z., Ammar, B., and Ayed, M. B. (2024). Towards au-

tomated optimization of residual convolutional neural

networks for electrocardiogram classiﬁcation. Cogni-

tive Computation, 16(3):1334–1344.

Honig, L. S. and Chin, S. S. (2001). Alzheimer’s dis-

ease. Science of Aging Knowledge Environment,

2001(1):dn2–dn2.

Madhat, M. I., Kadhim, K. N., Mohamed, F., Mohd Rahim,

M. S., Najjar, F. H., and Ramadhan, A. J. (2024). Di-

agnosing alzheimer’s disease severity: A comparative

study of deep learning algorithms. In BIO Web of Con-

ferences, volume 97, page 00102. EDP Sciences, EDP

Sciences.

Mokni, R. and Haoues, M. (2022). Cadnet157 model:

ﬁne-tuned resnet152 model for breast cancer diagno-

sis from mammography images. Neural Computing

and Applications, 34(24):22023–22046.

Rasmussen, J. and Langerman, H. (2019). Alzheimer’s dis-

ease - why we need early diagnosis. Degenerative

Neurological and Neuromuscular Disease, 9:123–

130. eCollection 2019.

Uraninjo (2024). Search results for alzheimer’s disease

augmentation datasets. https://www.kaggle.com/

datasets/uraninjo/augmented-alzheimer-mri-dataset.

Accessed: 2024-06-22.

Vengala, A. (2024). Classiﬁcation of mild, very mild,

moderate, and non-demented alzheimer’s disease mri

scans with svm.

DATA 2025 - 14th International Conference on Data Science, Technology and Applications

438