Deep Learning Approaches for Early Detection and Classification of

Alzheimer's Disease

Diqing Xu

School of Computer Science and Engineering & College of Software Engineering & School of Artificial Intelligence,

Southeast University, Nanjing, Jiangsu, China

Keywords: Deep Learning, Alzheimer, VGG16, Inception_v4, ResNet50.

Abstract: Alzheimer's Disease (AD) is a neurodegenerative condition that presents major obstacles to early diagnosis

and classification. This study proposes a new deep learning-based method to classify preprocessed brain MRI

scans, incorporating techniques for transfer learning and data augmentation. Three Convolutional Neural

Network (CNN) models were utilized: the 16-layer Visual Geometry Group network (VGG16), Inception

version 4 (Inception_v4), and the 50-layer Residual Network (ResNet50). The dataset used in this research,

sourced from Kaggle, contains around 6,400 MRI scans, categorized into four classes: mild dementia,

moderate dementia, non-demented, and very mild dementia. A tailored data augmentation pipeline was

developed, utilizing techniques such as rotation, flipping, and intensity modifications. This was combined

with transfer learning by employing pre-trained models from large-scale image datasets, which were then

fine-tuned for AD classification. The performance of the VGG16, Inception_v4, and ResNet50 models was

tested under four experimental scenarios: baseline (without augmentation or transfer learning), data

augmentation alone, transfer learning alone, and a combination of data augmentation and transfer learning.

The findings demonstrated that the integration of transfer learning and data augmentation substantially

enhanced classification accuracy, with the top-performing model achieving an accuracy of 98.49%. This

method can enhance the accuracy and reliability of AD diagnosis, contributing to more timely intervention

and treatment.

1 INTRODUCTION

Alzheimer's disease is a neurological condition that

gradually deteriorates cognitive and memory

functions., ultimately affecting the completion of

simple daily tasks. Currently, it is the seventh leading

cause of death in the United States, making early

diagnosis and intervention crucial (Smithsonian

Magazine, 2023).

Imaging techniques like Computed Tomography

(CT), Magnetic Resonance Imaging (MRI), and

Positron Emission Tomography (PET) are employed

in traditional diagnostic methods to diagnose

Alzheimer's disease (AD). While these methods assist

in diagnosis, they have limitations in accuracy and

cost. CT can detect obvious changes like brain

atrophy but has limited utility in early AD detection

(Adduru et al., 2020). MRI reveals detailed brain

structures, such as atrophy and hippocampal volume

https://orcid.org/0009-0007-1886-2003

reduction, but cannot directly identify amyloid

plaques and neurofibrillary tangles (Clark, 2003).

PET can image β-amyloid plaques and tau tangles but

is expensive, poses radiation risks, is primarily used

for research, and has limited long-term sensitivity

(Trudeau, 2018).

Deep learning and artificial intelligence have

made great progress in Alzheimer's disease diagnosis.

Convolutional Neural Networks (CNNs) can identify

brain atrophy patterns, particularly changes in the

hippocampus and entorhinal cortex (Subramoniam,

2022), which are important early features of AD. By

training on large datasets of labeled MRI images,

simplified CNN models can effectively differentiate

between healthy and diseased brains. El-Assy et al.

suggest that CNN models with simple structures can

achieve 95% accuracy in a five-classification task

(El-Assy, 2024).

340

Xu and D.

Deep Learning Approaches for Early Detection and Classiﬁcation of Alzheimer’s Disease.

DOI: 10.5220/0013517400004619

In Proceedings of the 2nd International Conference on Data Analysis and Machine Learning (DAML 2024), pages 340-348

ISBN: 978-989-758-754-2

By assessing model performance across four

cognitive states, this study seeks to increase the

accuracy of early Alzheimer's disease detection by

offering a comprehensive overview of the model's

diagnostic skills at various phases. This research will

provide important evidence for developing early

intervention and treatment strategies for Alzheimer's

disease, thereby improving patients' quality of life.

2 METHODOLOGY

2.1 Dataset

The Alzheimer's Disease MRI dataset that is freely

accessible on the Kaggle platform was used in this

investigation. Images from four categories—Non-

Demented, Very Mild Demented, Mild Demented,

and Moderate Demented—are included in this

dataset. The dataset is compiled from multiple

sources, including various websites, hospitals, and

public databases.

The collection provides approximately 6,400 2D

image slices with an original resolution of 208 x 176

pixels. These images represent patients across various

age groups and genders, as illustrated in Figure 1. For

uniformity and processing efficiency, all images have

been resized to 128 x 128 pixels.

Figure 1: Alzheimer's disease MRI dataset (Photo/Picture

credit: Original).

Figure 1 displays carefully selected slice planes

from comprehensive 3D scan datasets, stored in PNG

format to ensure superior image quality and retain

fine details. All images have undergone extensive

preprocessing to remove any identifiable information,

ensuring strict adherence to ethical guidelines and

protecting patient confidentiality.

2.2 Experimental Setup

The high-performance computer environment used

for all of the study's experiments allowed for the most

effective training and assessment of the deep learning

models. Table 1 provides a summary of all the

experimental setup's detailed specifications.

Table 1: Experimental environment configuration

Name Confi

uration Information

eratin

stem Ubuntu 22.04

Programming

Language

Python 3.10

Framewor

Torch 2.1.0

CUDA 12.1

CPU Intel

(

)

Xeon

(

)

Platinum 8362

GPU RTX 3090(24GB)

The integration of hardware and software

configurations established an optimal computing

environment for the experiments, enabling efficient

model training, optimization, and evaluation.

Moreover, it ensured the consistency and

reproducibility of the experimental results.

2.3 Data Preprocessing

2.3.1 Dataset Partitioning

The original dataset was split into training and

validation sets using a random sample strategy to

ensure the efficacy and generalizability of the models.

The split ratio of 70:30, which is frequently utilized

in deep learning and machine learning studies, was

utilized. This ratio ensures an ample amount of

training data while preserving a sufficient number of

validation samples for model assessment. The

detailed dataset distribution is as follows:

Training set: 4,479 images (approximately 70%)

Validation set: 1,919 images (approximately

30%)

The dataset comprises a total of 6,398 images.

This splitting strategy ensures that the training set

includes a sufficient number of samples to capture

complex feature patterns, while the validation set

remains large enough to accurately assess model

performance and detect potential overfitting.A

stratified random sampling technique was employed

to preserve the original class distribution (Non-

Demented, Very Mild Demented, Mild Demented,

and Moderate Demented) across both the training and

validation sets. This method reduces sampling bias

and enhances the model's stability across all

categories.

To evaluate the model's capacity for

generalization further, an independent test set was

reserved. This test set remained completely unused

during the training and tuning processes and was

exclusively applied for the final model evaluation.

Detailed information about this test set will be

provided in subsequent sections.

Deep Learning Approaches for Early Detection and Classiﬁcation of Alzheimer’s Disease

341

2.3.2 Data Augmentation

Extensive data augmentation techniques were used on

the training set to improve the model's generalization

and lower the possibility of overfitting. These

strategies were implemented using the argumentation

library, as detailed in Table 2. For the training set,

multiple augmentation methods were applied to

diversify the data.

In contrast, for the validation set, only center

cropping and standardization to maintain the original

characteristics of the data. This approach ensures that

the validation process accurately reflects the model's

output with unknown data, without the influence of

additional data augmentation. This carefully designed

data preparation and augmentation strategy, it aims to

maximize the value of available data while ensuring

the fairness and reliability of model evaluation.

Table 2: Data augmentation techniques

Feature Training set Validation set

Number of

Images

4479 1919

Proportion 70% 30%

Center Cro

Yes

(

128x128

)

Yes

(

128x128

)

Random

Horizontal Flip

Yes (50%

robability)

Color Jitte

Yes (±10%) No

Random Rotation Yes (-15° to

15°)

Random

Translation and

Scaling

Yes (max 5%) No

Normalization Yes Yes

2.3.3 Data Normalization

Using the mean values (0.485, 0.456, 0.406) and

standard deviations (0.229, 0.224, 0.225) calculated

from the ImageNet dataset, all images in this study

were normalized. These values are widely used in the

field of computer vision for transfer learning tasks.

The normalization process helps the models in this

experiment converge faster and improves their

generalization ability, while also ensuring

compatibility with pre-trained models.

2.3.4 Class Balancing

To ensure that every class is fairly represented in the

training set, oversampling was applied to the minority

classes, especially for the "Moderate Dementia"

class, which was significantly oversampled. This

ensured that the model could sufficiently learn the

characteristics of each class during training, thus

improving overall model performance and

generalization ability. Specifically, the "Moderate

Dementia" class was augmented 100 times, while the

"Non-Demented" class was only augmented twice.

Following the process of class balancing, the dataset

comprised 8,960 photos categorized as Non-

Demented, 6,663 images classified as Very Mildly

Demented, 8,778 images classified as Mildly

Demented, and 8,800 images classified as Moderately

Demented.

2.4 Transfer Learning Strategy and

Early Stopping Mechanism

This study conducted comparative experiments on

transfer learning and non-transfer learning using three

deep learning models: ResNet50, VGG16, and

Inception_v4. The training process was further

optimized by incorporating an early stopping

mechanism. The following is a comprehensive

analysis of the experimental results.

2.4.1 Advantages of Transfer Learning

Transfer learning greatly sped up the training process

and enhanced the model's performance on the

validation set by using pre-trained model weights

from massive datasets like ImageNet. Transfer

learning showed definite benefits for all three models:

ResNet50, VGG16, and Inception_v4.

Specifically, the training loss and validation loss of

the transfer learning models decreased rapidly, and

the validation accuracy reached a high level within

the early epochs (see Figures 2-3). This rapid

convergence indicates that the pre-trained features

provided by transfer learning effectively guided the

model’s optimization direction early on, allowing the

models to adapt to the target task more quickly and

accurately.

DAML 2024 - International Conference on Data Analysis and Machine Learning

342

(a) (b)

Figure 2: Inception_v4 - non-transfer learning loss and accuracy curves, (a) non-transfer learning loss results; (b) non-transfer

learning accuracy curve results (Photo/Picture credit: Original).

(a) (b)

Figure 3: Inception_v4 - transfer learning loss and accuracy curves, (a) Transfer learning loss results; (b) Transfer learning

accuracy curve results. (Photo/Picture credit: Original).

Figure 2 shows the trends in loss and accuracy for

the model without transfer learning (No Transfer

Learning). In (a), the training loss gradually decreases

with the increasing number of epochs and eventually

stabilizes, indicating that the model's performance on

the training set is progressively improving. However,

the validation loss, after an initial decline, stabilizes

and exhibits some fluctuations. This implies that the

model can be overfitting if it exhibits strong

performance on the training set but poor

generalization on the validation set.

The accuracy curves in (b) further confirm this.

The training accuracy increases significantly with

more epochs, eventually approaching 0.7, while the

validation accuracy, after a rapid initial rise,

fluctuates and ultimately settles around 0.6. This

suggests that perhaps as a result of the model's poor

generalization capacity, the model's performance on

the validation set is not as strong as it was on the

training set.

In contrast, Figure 3 shows the training results of

the model under transfer learning. In (a), the training

loss rapidly decreases and eventually approaches zero,

while the validation loss also exhibits a significant

downward trend. Although there is slight fluctuation

in the later stages, the overall validation loss remains

at a relatively low level. This suggests that transfer

learning improves the model's performance on the

validation set and successfully addresses the

overfitting problem.

The accuracy curves in (b) further support this

conclusion. In comparison to the non-transfer

learning model, the training accuracy increases

significantly under transfer learning and finally

approaches 0.9, while the validation accuracy

stabilizes around 0.8. This demonstrates that transfer

learning greatly improves the model's generalization

ability, enabling it to achieve higher accuracy on the

validation set.

Similarly, in the ResNet50 model, the validation

accuracy of the model using transfer learning

ultimately reached approximately 0.95, whereas the

model without transfer learning exhibited greater

fluctuations in validation accuracy and was unable to

Deep Learning Approaches for Early Detection and Classiﬁcation of Alzheimer’s Disease

343

achieve the same level. Both the VGG16 and

Inception_v4 models showed similar trends, where

the models with transfer learning demonstrated faster

decreases in training and validation losses, along with

significantly higher accuracy on the validation set.

2.4.2 Early Stopping Mechanism

The early stopping mechanism played a critical role

in the training process of all models, especially when

the performance on the validation set plateaued or

began to deteriorate. By setting a patience parameter,

the early stopping mechanism was able to terminate

training when no further improvements were

observed on the validation set, thus preserving the

model with the highest performance on the validation

set and preventing overfitting.

During the training of all transfer learning models,

the early stopping mechanism effectively reduced

unnecessary training time, ensuring that training was

halted once the model reached its optimal

performance. This approach not only improved

training efficiency but also sustained optimal model

performance on the validation set. For example, the

ResNet50 and Inception_v4 transfer learning models

ceased training when validation accuracy reached

approximately 0.95 and 0.85, respectively, while

validation loss remained consistently low.

2.4.3 Performance Comparison of Different

Models

While ResNet50, VGG16, and Inception_v4 have

distinct architectures, all three models showed

marked performance improvements when transfer

learning was paired with the early stopping

mechanism. However, differences in their

performance on specific tasks became evident. For

example, ResNet50 outperformed both VGG16 and

Inception_v4 on the validation set, likely due to the

strengths of its residual network in deep feature

extraction.

VGG16 displayed greater instability in non-

transfer learning conditions, resulting in lower

validation accuracy. This indicates that deeper

models may face challenges in learning effective

features from small datasets without pre-trained

weights. In contrast, Inception_v4 demonstrated

greater adaptability, achieving high accuracy

relatively quickly with transfer learning.

The performance and training efficiency of deep

learning models were significantly increased by the

combination of early stopping and transfer learning.

Models can quickly adjust to new tasks using transfer

learning with little data, and early halting helps avoid

overfitting. All models benefited from this approach,

with ResNet50 achieving the highest validation

accuracy. Future research could explore the

performance of various architectures in transfer

learning or further optimize the early stopping

mechanism to better meet diverse task requirements.

2.5 Comparison During the Training

Process

The experiment employed 4,479 training images and

1,919 validation images. The models' performance

was assessed by contrasting VGG16, Inception_v4,

and ResNet50 under four distinct experimental

conditions: baseline (no data augmentation or transfer

learning), data augmentation only, transfer learning

only, and a combination of both data augmentation

and transfer learning.

2.5.1 Validation Accuracy Under Different

Experimental Conditions

Table 3: Accuracy comparison of different models under

various experimental conditions

Experimental

Condition

VGG16 Inception_v4 ResNet50

Baseline (No

Data

Augmentation

or Transfer

Learning)

0.5020 0.5800 0.5690

Data

Augmentation

Only

0.5477 0.6071 0.6154

Transfer

Learnin

Onl

0.8800 0.8300 0.9050

Data

Augmentation

+ Transfer

Learnin

0.9823 0.8551 0.9922

Based on four experimental settings (Basis, Data

Augmentation Only, Transfer Learning Only, and

Data Augmentation + Transfer Learning), the

validation accuracy of the VGG16, Inception_v4, and

ResNet50 models was compared (Table 3). The

results show that the integration of transfer learning

and data augmentation yielded the maximum

precision for all models, with ResNet50 performing

best, achieving a validation accuracy of 0.9922.

VGG16 also demonstrated significant improvement,

reaching an accuracy of 0.9823 under this condition,

largely attributed to the augmentation techniques.

Inception_v4, though showing improvement,

experienced a relatively smaller gain, reaching an

accuracy of 0.8551 with the combined approach.

DAML 2024 - International Conference on Data Analysis and Machine Learning

344

2.5.2 Comparison of Loss Values Under

Different Experimental Conditions

Table 4: Loss values under four experimental conditions

Experimental

Condition

VGG16 Inception_v4 ResNet50

Baseline (No

Data

Augmentation

or Transfer

Learning)

1.0000 0.9500 0.9000

Data

Augmentation

Only

0.8804 0.8092 0.7841

Transfer

Learning Onl

0.5000 0.4500 0.3500

Data

Augmentation

+ Transfer

Learnin

0.1322 0.4073 0.0447

The trend in the loss values is consistent with the

validation accuracy results. When combining data

augmentation and transfer learning, the loss values

for all models decreased significantly, as shown in

Table 4. ResNet50 achieved the lowest loss value

under this condition, at 0.0447. VGG16's loss value

also dropped considerably to 0.1322. In comparison,

Inception_v4's loss value decreased to 0.4073, but the

improvement was not as pronounced as in the other

models.

2.5.3 Comparison of Training Time Under

Different Experimental Conditions

Table 5: Training time under four experimental conditions

Experimental

Condition

VGG16 Inception_v4 ResNet50

Baseline (No

Data

Augmentation

or Transfer

Learnin

10m 0s 12m 30s 6m 45s

Data

Augmentation

Only

12m

23s

15m 24s 8m 10s

Transfer

Learnin

Onl

20m

15s

22m 40s 5m 30s

Data

Augmentation

+ Transfer

Learning

25m 1s 29m 17s 7m 2s

In terms of training time, Table 5 illustrates how

all models' training times rose with the introduction

of data augmentation and transfer learning. ResNet50

had the shortest training time under all conditions,

particularly in the combined condition, where it took

only 7 minutes and 2 seconds. VGG16 and

Inception_v4 had relatively longer training times,

with 25 minutes and 1 second and 29 minutes and 17

seconds, respectively, under the combined condition.

3 RESULT

3.1 ROC Curve

Figure 4: ROC curves for the three models (Photo/Picture

credit: Original).

Through the analysis of the multi-class ROC

curves for the VGG16, Inception_v4, and ResNet50

models, this study found that the integration of

transfer learning with data augmentation significantly

improved the classification performance of the

Deep Learning Approaches for Early Detection and Classiﬁcation of Alzheimer’s Disease

345

models, as shown in Figure 4. Both VGG16 and

ResNet50 achieved an AUC of 1.00 across all

categories, demonstrating extremely high

classification accuracy and generalization ability,

making them suitable for complex task scenarios. In

contrast, Inception_v4 showed slight shortcomings in

some categories. Although its performance was

significantly enhanced by combining transfer

learning with data augmentation, the AUC values for

certain categories still did not reach the optimal level.

3.2 Confusion Matrix

Figure 5: Confusion matrix analysis of the three models

(Photo/Picture credit: Original).

Through the analysis of the confusion matrices, as

shown in Figure 5, it is evident that both VGG16 and

ResNet50 performed more consistently in the

classification task, particularly in distinguishing

between Class 2 and Class 3, where the

misclassification rate was significantly lower than

that of Inception_v4. Inception_v4 tended to struggle

when dealing with small differences between classes,

with a noticeable increase in misclassifications

between Class 2 and Class 3.

Overall, ResNet50 and VGG16 exhibited

excellent generalization and accuracy by utilizing a

blend of data enrichment and transfer instruction,

making them very successful for complex multi-class

classification tasks. By contrast, Inception_v4

exhibited a higher misclassification rate in certain

categories, revealing potential limitations in handling

more challenging or nuanced tasks.

3.3 Accuracy

Table 6: Comparison of accuracy under four experimental

conditions

Experimental

Condition

VGG16 Inception_v4 ResNet50

Baseline (No

Data

Augmentation

or Transfer

Learning)

0.5010 0.5504 0.6001

Data

Augmentation

Onl

0.5753 0.6326 0.6274

Transfer

Learning Onl

0.8354 0.7185 0.9592

Data

Augmentation

+ Transfer

Learning

0.9781 0.7353 0.9849

Table 6 compares the accuracy results of VGG16,

Inception_v4, and ResNet50 across four experimental

conditions. ResNet50 consistently outperformed the

other models in all scenarios, especially when both

transfer learning and data augmentation were applied,

achieving an accuracy of 0.9849. VGG16 also

showed notable improvement, reaching 0.9781 under

the same conditions. In contrast, despite the

enhancements from data augmentation,

Inception_v4’s highest accuracy reached only

0.7353, even when combined with transfer learning.

DAML 2024 - International Conference on Data Analysis and Machine Learning

346

3.4 Cohen's Kappa Coefficient

Table 7: Comparison of Cohen's kappa coefficients under

four experimental conditions

Experimental

Condition

VGG16 Inception_v4 ResNet50

Baseline (No

Data

Augmentation

or Transfer

Learning)

0.3507 0.3834 0.438

Data

Augmentation

Only

0.3034 0.4203 0.4169

Transfer

Learning Onl

0.5512 0.5816 0.696

Data

Augmentation

+ Transfer

Learnin

0.9639 0.5686 0.9752

When combining data augmentation with transfer

learning, ResNet50 achieved a precision of 0.9849

and a coefficient of Cohen's Kappa of 0.9752,

outperforming the other models (Table 7). This aligns

with Hasanah et al.'s findings, which also highlighted

the strong performance of ResNet in medical image

classification (Hasanah, 2023). However, ResNet50’s

longer training time, likely due to its deeper

architecture, was noted compared to simpler models.

VGG16 also showed notable improvement under

similar conditions, reaching a precision of 0.9781 and

a Cohen's Kappa coefficient of 0.9639, reflecting its

robust feature extraction capabilities, particularly

with transfer learning. However, the model's

extended training time could limit its applicability in

scenarios requiring rapid processing or real-time

updates.

Inception_v4, on the other hand, underperformed

compared to ResNet50 and VGG16, both in accuracy

and Cohen’s Kappa. Its highest accuracy was 0.7353,

with a Cohen's Kappa of 0.5686, likely due to the

complexity of its architecture, which made it less

efficient for MRI image analysis. Neshat et al.

similarly suggested that the Inception architecture

may require additional fine-tuning for certain medical

imaging tasks (Neshat, 2023). ResNet50 could

benefit from the incorporation of attention

mechanisms, which have been shown to enhance

model focus in medical image classification (Zhou et

al., 2022). For VGG16, knowledge distillation could

be a potential solution to its longer training time.

Hinton et al. demonstrated that this technique

effectively transfers knowledge from larger models to

smaller, faster ones, reducing training time while

maintaining accuracy (Hinton, 2015).

One limitation of this study is the diversity of the

dataset. Although approximately 6,400 MRI images

were included, this may not be enough to represent all

possible cases and variations. Future studies should

aim to expand both the diversity and size of the

dataset to better assess model performance across

various populations and Alzheimer's disease stages.

Additionally, this study primarily focused on 2D MRI

images, which may have overlooked crucial spatial

information available in 3D MRI data, potentially

impacting diagnostic accuracy.

Based on these findings and limitations, several

future research directions are recommended. First,

expanding the dataset to include more diverse MRI

data from patients of varying ethnicities, age groups,

and disease stages would enhance the model’s

generalizability. Second, the potential of 3D CNN

models should be explored, as Folego’s research

showed that 3D CNNs offer significant advantages in

diagnosing brain diseases (Folego, 2020).

Incorporating 3D MRI data could improve the

precision of Alzheimer’s diagnosis. Lastly, exploring

multimodal data fusion is essential. Song et al. found

that combining data from PET scans, genetics, and

other sources significantly improves diagnostic

accuracy in neurodegenerative diseases (Song et al.,

2021). These advancements could enhance the

reliability of models for deep learning in early

Alzheimer’s detection and offer more robust support

for clinical practice.

4 CONCLUSIONS

To identify and classify AD early on, this paper

proposes a deep learning-based method based on

brain MRI data. Through the combination of data

augmentation and transfer learning, the study

demonstrated considerable gains in classification

accuracy by comparing three CNN architectures:

ResNet50, Inception_v4, and VGG16. ResNet50, the

best-performing model, achieved 98.85% accuracy,

surpassing the baseline. These findings highlight the

potential of this approach to improve AD

classification accuracy with broad application

potential.

The innovation of this study lies in integrating

customized data augmentation techniques with

transfer learning to efficiently utilize small MRI

datasets, excelling in neurodegenerative disease

classification. This approach not only improves AD

classification but also serves as a reference for other

medical imaging tasks.

Deep Learning Approaches for Early Detection and Classiﬁcation of Alzheimer’s Disease

347

Future research will focus on enhancing model

generalization and incorporating multimodal data

fusion. To improve generalization, the study aims to

collect multicenter data, standardize preprocessing

workflows, and apply domain adaptation techniques

to ensure robustness across clinical environments. For

multimodal data fusion, future efforts will integrate

biomarkers like MRI, PET scans, and cerebrospinal

fluid analysis, while developing architectures to

process heterogeneous data and explore optimal

fusion strategies. These efforts are expected to further

improve AD diagnosis and provide new insights into

diagnosing other neurodegenerative diseases, leading

to more accurate and comprehensive diagnostic

systems.

REFERENCES

Adduru, V., Baum, S. A., Zhang, C., Helguera, M., Zand,

R., Lichtenstein, M., Griessenauer, C. J. and Michael,

A. M., 2020. A method to estimate brain volume from

head CT images and application to detect brain atrophy

in Alzheimer disease. American Journal of

Neuroradiology, 41(2), pp.224-230.

Clark, C. M., Xie, S., Chittams, J., Ewbank, D., Peskind, E.,

Galasko, D., Morris, J. C., McKeel, D. W., Farlow, M.,

Weitlauf, S. L., Quinn, J., Kaye, J., Knopman, D., Arai,

H., Doody, R. S., DeCarli, C., Leight, S., Lee, V. M.-Y.

and Trojanowski, J. Q., 2003. Cerebrospinal fluid tau

and β-amyloid: How well do these biomarkers reflect

autopsy-confirmed dementia diagnoses? Archives of

Neurology, 60(12), pp.1696–1702.

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

Mohamed, M. A., 2024. A novel CNN architecture for

accurate early detection and classification of

Alzheimer’s disease using MRI data. Scientific

Reports, 14(1), 3463.

Folego, G., Weiler, M., Casseb, R. F., Pires, R. and Rocha,

A., 2020. Alzheimer's disease detection through whole-

brain 3D-CNN MRI. Frontiers in Bioengineering and

Biotechnology, 8, 534592.

Hasanah, S. A., Pravitasari, A. A., Abdullah, A. S., Yulita,

I. N. and Asnawi, M. H., 2023. A deep learning review

of ResNet architecture for lung disease identification in

CXR images. Applied Sciences, 13(24), 13111.

Hinton, G., Vinyals, O. and Dean, J., 2015. Distilling the

knowledge in a neural network. arXiv.

Neshat, M., Ahmed, M., Askari, H., Thilakaratne, M. and

Mirjalili, S., 2023. Hybrid Inception Architecture with

residual connection: Fine-tuned Inception-ResNet deep

learning model for lung inflammation diagnosis from

chest radiographs. arXiv.

Smithsonian Magazine, 2023. Here’s where the highest

rates of Alzheimers are in the United States.

Song, J., Zheng, J., Li, P., Lu, X., Zhu, G. and Shen, P.,

2021. An effective multimodal image fusion method

using MRI and PET for Alzheimer's disease diagnosis.

Frontiers in Digital Health, 3, 637386.

Subramoniam, M., Aparna, T. R., Anurenjan, P. R. and

Sreeni, K. G., 2022. Deep learning-based prediction of

Alzheimer’s disease from magnetic resonance images.

In M. Saraswat, H. Sharma and K. V. Arya (Eds.),

Intelligent vision in healthcare, pp.183-198. Springer.

Trudeau, V. L., 2018. Facing the challenges of

neuropeptide gene knockouts: Why do they not inhibit

reproduction in adult teleost fish? Frontiers in

Neuroscience, 12, Article 302.

Zhou, Z., Lu, C., Wang, W., Dang, W. and Gong, K., 2022.

Semi-supervised medical image classification based on

attention and intrinsic features of samples. Applied

Sciences, 12(13), 6726.

DAML 2024 - International Conference on Data Analysis and Machine Learning

348