Dementia Identification using a Class of CNN based Methods and

Transfer Learning

LIZA MEDHI, BEAUTI PRIYA CHOUDHURY, SURAJIT DEKA, KANDARPA KUMAR SARMA

Department of Electronics and Communication Engineering,

Gauhati University,

Guwahati-781014, Assam,

INDIA

Abstract: - Through the use of transfer learning techniques and multiple Convolutional Neural Networks (CNNs),

our study offers a novel method for the early identification of dementia. Through data augmentation, we ensure

robustness and enhance the model's ability to extract complex patterns from MRI data by utilizing pre-trained

models such as VGG19, ResNet50, and Inception V3. We improve the identification of dementia-related patterns

by utilizing well-established models, presenting promising advances in our field. In addition, our study explores the

theoretical underpinnings of using MRI imaging to differentiate between different phases of dementia, offering

important new perspectives on the course of the disease. We initialized the models with weights of pre-trained

image net equivalents and perform dataset pre-processing, segmentation. Our results well exceeds with VGG19

topping with 91.5% accuracy followed by Inception V3 as 87.19%. ResNet50 also got an impressive score of

74.76%. The future work shows the potential to identify clearer differences to be used as early diagnostic aid for

Dementia in patients who don’t have neurological symptoms with mere CNNs using transfer learning.

Key-Words: - Cognitively Normal (CN), Convolutional Neural Network (CNN), Deep Neural Network (DNN),

Mild Cognitive Impairment (MCI), Residual Neural Network (ResNet), Visual Geometry Group

(VGG).

Received: March 14, 2024. Revised: October 15, 2024. Accepted: November 13, 2024. Published: December 31, 2024.

1 Introduction

Dementia, is a disease of brain which is irreversible. It

affects one’s memory, thinking capability and many

other. It is a long-term neurodegenerative disease and

is related to one person’s loss of memory which

affects the decision-making skill of the person and

also can obstruct social activity, [1]. Therefore, early

detection of the disease is very important for the

patient’s health.

In the United States, it is one of the most

important cause for death having 3.6% of all

casualties in the 2014 death statistics, [2]. Artificial

intelligence (AI) based tools are preferred to provide

preliminary diagnostic support in case of many

disease. Likewise, AI tools have also been used for

the early detection of the condition of a patient

suffering from dementia. Many AI tools can be

effectively design and configured to provide initial

support to the patients and the healthcare community

to formulate treatment for people suffering from this

disorder. Through this research we aim to use CNN

and Transfer learning by combining deep learning

techniques with the existing knowledge to thoroughly

study the MRI scans. Using well-known pre-trained

models like VGG-19, Inception V3 and ResNet-50

will help us to make some progress in this field, [3].

These models act like tools that identify important

patterns from the MRI scans, which are useful for

detecting dementia. This helps the system to learn

faster and accurately detect the disease.

Transfer learning makes the task of applying

sophisticated models to brain scans easier by

customizing pre-existing neural networks to address

the unique difficulties associated with dementia

detection, [4].

1.1 Motivation

Detection of dementia beforehand is an important

element in the process of taking corrective steps to

arrest the onset of the serious stages of the disease. If

the early symptoms of the disease are detected early,

treatment can be initiated to arrest the degradation of

WSEAS TRANSACTIONS on COMPUTERS

DOI: 10.37394/23205.2024.23.24

Liza Medhi, Beauti Priya Choudhury,

Surajit Deka, Kandarpa Kumar Sarma

E-ISSN: 2224-2872

245

Volume 23, 2024

the condition of the patient. Once dementia sets in, a

person’s ability to carry out day to day functioning is

seriously affected, [1]. Several serious difficulties like

inability to concentrate, errors in judgment, forgetting,

repeating tasks, or saying the same sentences etc. are

observed. At times, these might prove to be life-

threatening for the patient. Each year the world-wide

death rates due to dementia are rising. Many deaths

due to dementia remain unreported, [5]. The

seriousness of the situation and the requirement to

provide respite to patients suffering from such

disorders have become a necessity. One of the key

attributes in this regard is to formulate technology

supported solutions. Such support through which early

detection of onset of dementia is possible become

critical to help a section of the society which are likely

to be affected by this disease, [6]. If such solutions

provide early detection, personalized care can be

initiated. While the onset of such a disease is

ascertained by the use of magnetic resonance imaging

(MRI), trained healthcare professionals give the

decision regarding the stages of the disorder. Lately, it

has been observed that several AI based approaches

have been adopted for detection of the onset of

dementia and its stages from MRI samples. MRI

samples applied to train CNN are found to be

effective for such a purpose. In this work we highlight

the application of a set of CNN based methods for

detection of dementia.

1.2 Contribution

The key contributions of our works can be

summarized as below:

 Firstly, a hybrid method formed using transfer

learning and the CNN is designed, configured and

trained for detecting dementia using MRI

samples. After extensive training and testing we

find that the combination is robust and reliable

while providing reliable decision regarding the

occurrence of dementia.

 Next, we apply certain data augmentation

methods to increase the number of training

samples and to nullify the ill-effects of having

insufficient MRI samples. As a result, the training

becomes robust and the method turns out to be

reliable,

 Finally, several pre-trained models like Inception

V3, ResNet50, and VGG-19 are employed for the

purpose to ascertain the effectiveness of these AI

tools.

2 Background Study

AI-based tests showcased the ability to improve the

detection of dementia process, according to past

research. These tests give physicians more precise and

reliable findings than older methods, and they provide

a direct, unbiased sign of dementia in its early stages,

[7]. Robust CNN models have shown great diagnostic

accuracy in dementia diagnosis after being trained and

evaluated on multiple independent groups, [8].

Relevance maps, which correlate hippocampal

relevance scores with volume - a significant MRI

marker for Alzheimer's disease - were used to assess

these models, [8]. Bayesian networks for Alzheimer's

disease detection have been developed, tested, and

examined with machine learning techniques, [9].

These networks derive interpretation and meaning

from complex data using probability models. These

become more reliable with each training cycle. This is

true with early detection of dementia as well, [9]. The

work [10] reported the application of a set of

techniques to distinguish between various groups of

speakers categorized as healthy, with mild cognitive

impairments (MCI), and mild Alzheimer’s disease.

Moreover, the work [3] reported the formulation of

techniques that distinguishes signs and symptoms of

MCI and Alzheimer’s disease applying distinct

patterns and markers which are linked with the

disorder. This has enhanced the accuracy of the

clinical analysis.

3 Proposed Approach

In this work, we adopted and trained three popular

CNN models namely VGG19, ResNet50, and

Inception V3 for identification of dementia using MRI

samples. The VGG19 is efficient in feature learning

and image identification tasks due to unique 19 layers

constituted by 3x3 convolutional filters, [11].

Additionally, the VGG19 is an excellent candidate for

applying transfer learning in specialized neuroimaging

requirements for deriving critical decisions about

dementia onset using MRI samples. VGG19 is known

to be efficient in recognizing complex patterns and

structure engrossed in MRI images, [12].

Residual Network has 50 layers hence called

ResNet50, is configured using a distinctive

architecture where there are residual blocks that are

trained to handle the issue of vanishing gradients, [3].

The network is known for its efficiency and has been

found to be excellent for image classification.

WSEAS TRANSACTIONS on COMPUTERS

DOI: 10.37394/23205.2024.23.24

Liza Medhi, Beauti Priya Choudhury,

Surajit Deka, Kandarpa Kumar Sarma

E-ISSN: 2224-2872

246

Volume 23, 2024

ResNet50's residual connections contribute towards its

neuroimaging efficiency by strengthening its ability to

capture subtle variations in MRI scans. It improves

the accuracy of the dementia detection, [3].

The Inception V3 is constituted by the use of

mixed convolutional filters in one layer and

processing in the subsequent layers through

simultaneous flow of the input streams. Also, the

computation efficiency of the Inception V3 is

renowned [3] which makes it a reliable candidate for

analysing MRI data and recognizing early signs of

dementia.

While the residual blocks in ResNet50 helps in

preventing the occurrence of the vanishing gradient

problem, the Inception V3 provides computational

efficiency and the VGG19 demonstrates its excellence

in capturing minute variations in the MRI data, these

networks form a reliable platform for efficient

discrimination of MRI samples for reliable detection

of onset of dementia, [12]. A wide variety of features

can be gathered with great accuracy using Inception

V3's mixed filter system. These models together will

allow us to further enhance dementia detection

accuracy and progress neuroimaging methods.

4 Proposed Method

4.1 Configuring and Training the CNN for the

Proposed Approach

We used three well-known architectures—VGG19,

Inception V3, and ResNet50—for CNN models to

detect dementia. For better performance, transfer

learning was applied to initialize these models with

weights that had already been learned on the

ImageNet dataset, [12]. In order to improve model

interoperability, the MRI images were adjusted to

match the input dimensions used during ImageNet

training. The deep design and 3x3 filters of VGG19

allowed it to pick up intricate elements in the structure

of the brain, [12]. Global average pooling eliminated

overfitting while facilitating effective feature

extraction via Inception V3's novel convolutional

connections, [12]. Training efficiency was boosted by

ResNet50's application of remnant connections,

paving the way for easier convergence. We improved

the pre-trained models capacity to identify

dementia stages by fine-tuning them employing

labeled MRI data, making them useful tools for

identifying the minute abnormalities in the brain.

4.2 Block Diagram

The process logic of the proposed work is summarized

using Figure 1.

Fig. 1: Block diagram

In this research, we used three CNN models-

VGG19, Inception V3, and ResNet50, to use transfer

learning to the identification of different stages of

dementia based on neuroimaging data. We were able

to improve model performance and accelerate

convergence for the dementia detection challenge by

using pre-trained ImageNet weights. The models were

able to effectively record crucial brain properties

including structural shifts and changes in volume,

which are important signs of dementia, because the

MRI images had been pre-processed and shrunk to

meet the necessary input dimensions.

After feature extraction, a dropout was applied to

the data before it was sent through fully connected

layers to prevent overfitting, [12]. A softmax layer

was used for the final classification, giving each

dementia form its probability. utilizing an accurate

categorization approach, the dataset was split into four

categories: non-dementia, mild dementia, moderate

dementia, and very mild dementia. With the help of

each model's unique strengths, pattern recognition in

Inception V3, deeper learning in ResNet50, and broad

feature selection in VGG19 the phases of dementia

could be successfully classified, (Figure 2).

WSEAS TRANSACTIONS on COMPUTERS

DOI: 10.37394/23205.2024.23.24

Liza Medhi, Beauti Priya Choudhury,

Surajit Deka, Kandarpa Kumar Sarma

E-ISSN: 2224-2872

247

Volume 23, 2024

Fig. 2: Classification using CNN models

5 Results and Analysis

In this section we discuss the results and also include

certain discussion. The results are in terms of the

performance obtained from three different networks.

Learning curves obtained during training and

validation checks are shown. Further, confusion

matrix is used to depict the cases of different stages of

dementia.

5.1 VGG 19 Model

Fig. 3: VGG 19 Model accuracy

Figure 3 shows that the model achieved a

validation accuracy of 91.50%.

Fig. 4: VGG 19 MODEL LOSS

Figure 4 shows that the model achieved a

validation loss of 21.99%.

Fig. 5: VGG 19 Model confusion matrix

In Figure 5, the confusion matrix shows the true

positive counts for dementia diagnosis across different

stages: 625 for Non-Demented, 635 for Very Mild

Demented, 577 for Mild Demented, and 510 for

Moderate Demented.

WSEAS TRANSACTIONS on COMPUTERS

DOI: 10.37394/23205.2024.23.24

Liza Medhi, Beauti Priya Choudhury,

Surajit Deka, Kandarpa Kumar Sarma

E-ISSN: 2224-2872

248

Volume 23, 2024

5.2 ResNet-50 Model

Fig. 6: ResNet-50 Model accuracy

Figure 6 shows that the model achieved a

validation accuracy of 74.76%.

Fig. 7: ResNet-50 Model loss

Figure 7 shows that the model achieved a

validation loss of 60%.

5.3 Inception V3 Model

Fig. 8: Inception V3 Model accuracy

Figure 8 shows that the model achieved a

validation accuracy of 88.18%.

Fig. 9: Inception V3 Model loss

Figure 9 shows that the model achieved a

validation loss of 32.70%.

Therefore, the highest accuracy, at 91.50%, was

achieved with VGG19, underscoring its effectiveness

in dementia detection.

WSEAS TRANSACTIONS on COMPUTERS

DOI: 10.37394/23205.2024.23.24

Liza Medhi, Beauti Priya Choudhury,

Surajit Deka, Kandarpa Kumar Sarma

E-ISSN: 2224-2872

249

Volume 23, 2024

5.4 Prediction of MRI Images

Fig. 10: Prediction of MRI images

Fig. 10 shows predicted images of different stages

of dementia obtained from trained VGG19. There are

cases of very mild dementia, mild dementia, moderate

dementia and non-dementia cases. The MRI images

collected from public database and labelled using

expert knowledge are used to train the deep networks

and obtain the predicted state of the dementia.

6 Conclusion

In conclusion, our study focused on the importance of

identifying brain biomarkers from MRI scans in order

to recognize early warning signs of dementia. We

used three alternative models of deep neural networks,

each with unique architectural strengths: VGG19,

Inception V3, and ResNet50. After extensive pre-

processing, we used transfer learning to separate the

dataset into training, testing, and validation sets. We

then initialized our models using ImageNet's pre-

trained weights. With VGG19 scoring approximately

91.5% accuracy, Inception V3 scoring 87.19%, and

ResNet50 scoring 74.76%, our results indicated good

performance. These results highlight the benefit of

pre-trained models in the classification of dementia

and the future potential of neural networks for more

precise and timely diagnosis. Our results support

greater efforts to improve neurological illness

diagnosis, especially as dementia advances through

the use of neuroimaging datasets.

Although we classified genetic markers related to

dementia based on our analysis of MRI visuals, future

research will need access to OCT images of Dementia

patients, which are not yet available.

Declaration of Generative AI and AI-assisted

Technologies in the Writing Process

During the preparation of this work the authors used

Grammarly for language editing. After using this

service, the authors reviewed and edited the content as

needed and take full responsibility for the content of

the publication.

References:

[1] P. Saltz, S. Y. Lin, S. C. Cheng and D. Si,

"Dementia Detection using Transformer-Based

Deep Learning and Natural Language

Processing Models," 2021 IEEE 9th

International Conference on Healthcare

Informatics (ICHI), Victoria, BC, Canada, 2021,

pp. 509-510, doi:

10.1109/ICHI52183.2021.00094.

[2] Jiaquan Xu, Kenneth D Kochanek, Sherry L

Murphy, and Betzaida Tejada-Vera. Deaths:

final data for 2014. 2016.

[3] M. T. Abed, U. Fatema, S. A. Nabil, M. A.

Alam and M. T. Reza, "Alzheimer's Disease

Prediction Using Convolutional Neural Network

Models Leveraging Pre-existing Architecture

and Transfer Learning," 2020 Joint 9th

International Conference on Informatics,

Electronics & Vision (ICIEV) and 2020 4th

International Conference on Imaging, Vision &

Pattern Recognition (icIVPR), Kitakyushu,

WSEAS TRANSACTIONS on COMPUTERS

DOI: 10.37394/23205.2024.23.24

Liza Medhi, Beauti Priya Choudhury,

Surajit Deka, Kandarpa Kumar Sarma

E-ISSN: 2224-2872

250

Volume 23, 2024

Japan, 2020, pp. 1-6, doi:

10.1109/ICIEVicIVPR48672.2020.9306649

[4] R. Ribani and M. Marengoni, "A Survey of

Transfer Learning for Convolutional Neural

Networks," 2019 32nd SIBGRAPI Conference

on Graphics, Patterns and Images Tutorials

(SIBGRAPI-T), Rio de Janeiro, Brazil, 2019, pp.

47-57, doi: 10.1109/SIBGRAPI-T.2019.00010.

[5] Association Alzheimer’s. 2015 alzheimer’s

disease facts and figures. Alzheimer’s &

dementia: the journal of the Alzheimer’s

Association, 11(3):332, 2015, [Online].

https://www.alz.org/media/documents/2015facts

andfigures.pdf (Accessed Date: October 4,

2024).

[6] T. Subetha, R. Khilar and S. K. Sahoo, "An

Early Prediction and Detection of Alzheimer's

Disease: A Comparative Analysis on Various

Assistive Technologies," 2020 International

Conference on Computational Intelligence for

Smart Power System and Sustainable Energy

(CISPSSE), Keonjhar, India, 2020, pp. 1-5, doi:

10.1109/CISPSSE49931.2020.9212240

[7] R. Li, X. Wang, K. Lawler, S. Garg, Q. Bai, and

J. Alty, “Applications of artificial intelligence to

aid early detection of dementia: A scoping

review on current capabilities and future

directions,” Journal of Biomedical Informatics,

vol. 127, p. 104030, Mar. 2022, doi:

10.1016/j.jbi.2022.104030.

[8] M. Dyrba et al., “Improving 3D convolutional

neural network comprehensibility via interactive

visualization of relevance maps: evaluation in

Alzheimer’s disease,” Alzheimer’s Research &

Therapy, vol. 13, no. 1, Nov. 2021, doi:

10.1186/s13195-021-00924-2.

[9] W. H. Land and J. D. Schaffer, “A Machine

Intelligence Designed Bayesian Network

Applied to Alzheimer’s Detection Using

Demographics and Speech Data,” Procedia

Computer Science, vol. 95, pp. 168–174, 2016,

doi: https://doi.org/10.1016/j.procs.2016.09.308.

[10] G. Gosztolya, V. Vincze, L. Toth, M. Pakaski, J.

Kalman, and I. Hoffman, "Identifying Mild

Cognitive Impairment and mild Alzheimer’s

disease based on spontaneous speech using ASR

and linguistic features," Computer Speech &

Language, vol. 52, pp. 37-55, Aug. 2018, doi:

10.1016/j.csl.2018.07.007.

[11] A. Fathima, F. Taranum, M. Hijab, S. M. A.

Hashmi, S. S. Ahmad and G. Gupta,

"Classifying Alzheimer Disease using

VGG19," 2024 11th International Conference

on Computing for Sustainable Global

Development (INDIACom), New Delhi, India,

2024, pp. 1749-1753, doi:

10.23919/INDIACom61295.2024.10498804.

[12] Ian Goodfellow and Yoshua Bengio and Aaron

Courville, "Deep Learning" [Online]

https://www.deeplearningbook.org/ (Accessed

Date: October 4, 2024).

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.

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_

WSEAS TRANSACTIONS on COMPUTERS

DOI: 10.37394/23205.2024.23.24

Liza Medhi, Beauti Priya Choudhury,

Surajit Deka, Kandarpa Kumar Sarma

E-ISSN: 2224-2872

251

Volume 23, 2024