BenchCouncil Transactions on Benchmarks, Standards and

Evaluations, 2026

DOI: https://doi.org/10.66834/798kjf21

Research Article

RESEARCH ARTICLE

Design and Evaluation of an Interpretable Multimodal

Deep Learning Framework for Early Alzheimer’s

Disease Detection

Shehu Mohammed,

1,∗

Neha Malhotra

and Anmol Singh Rai

School of Computer Applications, Lovely Professional University, Phagwara, India and

Shrimann Superspeciality Hospitals, Jalandhar,

India

∗

Corresponding author. mohammedshehumafara@gmail.com

Received on 22 December 2025; Accepted on 24 June 2026

Abstract

Alzheimer’s disease is a progressive neurodegenerative disorder that signiﬁcantly impairs memory and cognitive functions

and aﬀects over 55 million people worldwide. The successful management and planning require early and accurate diagno-

sis. Conventional radiological assessment is often subjective and time-consuming, which highlights the need for automated

and reliable diagnostic solutions. Most deep learning models show promise for classifying neuroimaging data, but they tend

to be less computationally eﬃcient and less interpretable, and they cannot be integrated into patient-centric processes.

The gap between developing diagnostic algorithms with high accuracy and implementing them in a supportive framework

that includes patients and caregivers is large. This paper introduces a comprehensive, hybrid framework that addresses

these gaps. We present a dual-modality diagnostic system: a deep learning pipeline using EﬃcientNetV2-S for CT scan

classiﬁcation, complemented by a Feedforward Neural Network (FNN) that analyses structured clinical data for holistic

patient assessment. This diagnostic core is integrated into a user-friendly graphical user interface (GUI) and supplemented

by ”NeuroBot,” an AI-powered chatbot that provides domain-speciﬁc information and support. The two models have

been trained using the transfer learning method on a curated dataset of 30,000 brain CT slices. The EﬃcientNetV2-S

model achieved an accuracy of 98.19%. After hyperparameter tuning, the FNN model achieved an optimised accuracy

of 87.21%. The importance of the features addressed by the models was proved with the help of the statistical t-tests of

the corresponding clinical data. The integrated system enables a scalable, translatable, and patient-centered system to

improve the early analysis and treatment of Alzheimer’s disease.

Key words: Alzheimer’s Disease, CT Scan Classiﬁcation, EﬃcientNetV2-S, Deep Learning, AI Chatbot, Medical

Imaging, Transfer Learning

1. Introduction

Alzheimer’s disease (AD) is the most prevalent form of de-

mentia and primarily aﬀects the aging population, accounting

for approximately 60-70% of dementia cases worldwide. It is

a neurodegenerative, advanced, and progressive disorder that

slowly aﬀects the cognitive functions, starting with the mem-

ory and then with the dysfunction of judgment, language, and

behaviour. The disease has a signiﬁcant social impact, placing

substantial emotional and ﬁnancial burdens on patients, fam-

ilies, and healthcare systems worldwide. Early and accurate

diagnosis is therefore critical important. The timely diagnosis

will enable patients and the individuals who provide care to in-

vestigate the treatment interventions, develop useful long-term

care plans, and eventually improve the quality of life of patients

as the disease advances[1, 2].

The neuroimaging procedures are the core of the AD diag-

nostics process, and they allow clinicians to observe structural

changes of the brain typical of the disease, e.g., cortical at-

rophy and ventricular enlargement. Among these techniques,

Computed Tomography (CT) scans serve as a vital tool. CT

imaging is particularly valuable due to its rapid acquisition

time, widespread availability, and relative cost-eﬀectiveness

compared to other modalities like Magnetic Resonance Imaging

(MRI), making it a cornerstone of initial neurological workups,

especially in resource-limited settings[3]. However, the conven-

tional analysis of these scans is not without its challenges. The

diagnostic process is a subjective one and depends on the radiol-

ogist’s interpretation, and this cannot be adequately performed

in a short time, besides being prone to inter-rater inconsistency

and human error, especially in detecting the changes at the

initial stages of AD.

Shehu Mohammed et al.

These diagnostic limitations have been a primary driver for

the development of computer-aided diagnostic (CAD) systems,

a ﬁeld that has been revolutionized by the advent of deep learn-

ing. As extensively documented in systematic reviews, both

machine learning and deep learning techniques have demon-

strated profound success in the automated analysis of medical

images[4]. CNNs have been exceptionally skilled at this task,

in particular. Their architecture allows them to automatically

learn and identify complex, hierarchical patterns within visual

data, enabling the detection of subtle pathological indicators

in MRI and brain CT slices that may be imperceptible to the

human eye[5, 6]. This capability has spurred the creation of

numerous advanced models, including various hybridized deep

learning approaches, all aimed at pushing the boundaries of

diagnostic accuracy and reliability[7].

Despite this remarkable progress, a signiﬁcant gap per-

sists between the development of high-accuracy algorithms in

a research setting and their practical deployment in clinical

workﬂows. Many state-of-the-art CNN models face consider-

able hurdles, including the need for substantial computational

resources, which limits their real-time application. Moreover,

the inherent ”black-box” nature of many deep learning mod-

els can be a barrier to clinical adoption; for a diagnosis to be

trusted, clinicians must be able to understand and verify the

reasoning behind it, a critical factor in diﬀerential diagnosis[8].

Beyond the technical challenges, most existing research has fo-

cused almost exclusively on the diagnostic algorithm itself. This

narrow focus often neglects the broader clinical ecosystem and

the critical need for an integrated, user-friendly system that

not only provides a diagnosis but also supports patients and

caregivers with accessible, contextual information[9].

To overcome these complex issues, this paper proposes a

new and unique hybrid framework that combines the high-

performance diagnostic engine with an interactive patient sup-

port system that is based and comprehensive. Our primary

contribution is the development of a dual-model deep learning

pipeline that leverages a powerful and eﬃcient EﬃcientNetV2-S

architecture for the classiﬁcation of brain CT slices with ex-

ceptional accuracy and computational eﬃciency[10]. The core

novelty of our work lies in embedding these performant and

interpretable models, which utilize Grad-CAM for visual ex-

planations, within a complete, end-to-end ecosystem. This

system includes an intuitive graphical user interface (GUI) for

seamless interaction by clinicians and an AI-powered chatbot,

NeuroBot, designed to answer AD-related queries from pa-

tients and caregivers. By moving beyond mere classiﬁcation,

this holistic approach creates a practical, scalable, and sup-

portive tool designed to enhance early Alzheimer’s detection

and improve the overall standard of patient care[11]. Further-

more, to capture non-visual risk factors and cognitive metrics

that are crucial for a comprehensive diagnosis, our framework

integrates a secondary pathway that leverages a robust FNN to

analyze structured patient clinical records.

Although many deep learning models have achieved high

diagnostic accuracy for Alzheimer’s disease detection, most

studies focus primarily on algorithm development and eval-

uation using experimental datasets. These models are rarely

integrated into user-friendly systems that support clinical work-

ﬂows, patient interaction, or caregiver guidance. As a result,

there remains a gap between high-performance diagnostic al-

gorithms and practical systems that can be deployed in real

clinical environments. The proposed framework addresses this

gap by integrating the diagnostic models within an interac-

tive ecosystem that includes a graphical user interface and

an AI-based assistant to support clinicians, patients, and

caregivers.

The main contributions of this study can be listed as follows:

• Development of a dual-modality deep learning framework,

which uses EﬃcientNetV2-S for CT image classiﬁcation,

and a Feedforward Neural Network for structured clinical

data analysis.

• The application of explainable AI, which uses Grad-CAM

for region highlighting on brain CT slices, thereby improv-

ing their interpretability.

• Development of a user-centric deployment platform, which

includes a GUI and an AI-based chatbot, referred to as

NeuroBot.

• The validation of clinical features through statistical anal-

ysis, which includes independent sample t-tests and corre-

lation analysis to ascertain the signiﬁcance of cognitive and

lifestyle factors, such as MMSE and ADLs.

2. Literature Review

The ML and DL applied to the neurology practice have com-

pletely changed the Alzheimer Disease (AD) diagnostic process

and provided meaningful and data-intuitive answers to the cur-

rent practices. The evolution of these techniques has been

rapid, moving from foundational models to highly sophisti-

cated, specialized architectures. Early research in automated

AD detection was primarily centered on classical machine

learning algorithms. Models such as Support Vector Machines

(SVMs), Random Forests, and Decision Trees were applied to

neuroimaging data, but their eﬃcacy was often constrained by

a reliance on handcrafted feature extraction[12]. This process,

which required extensive domain knowledge to manually deﬁne

and extract relevant features like hippocampal volume or corti-

cal thickness, was not only labor-intensive but also limited the

models’ ability to discover novel, complex patterns within the

data. Despite the need to take such measures, the subsequent

emergence of deep learning, and, particularly, the Convolu-

tional Neural Networks (CNNs) became a game-changer in the

matter. The CNNs have revolutionised the study of medical

images; they enable end-to-end learning of hierarchical fea-

tures beneath the raw pixel data. This capability has led to

a proliferation of studies demonstrating the accurate predic-

tion and diagnosis of AD using a variety of deep learning

models, which consistently outperform their traditional ML

counterparts[13, 14].

The current research in the area, however, has predom-

inantly been interested in the application of deep learning

to analyze neuroimaging images, and, more speciﬁcally, in

Magnetic Resonance Imaging (MRI) and, more recently, in

Computed Tomography (CT) scans. The authors have experi-

mentally shown that applying advanced image processing and

enhancement techniques before model training improves AD de-

tection accuracy[15]. Recognizing that a single data source may

not capture the complexity of AD pathology, numerous stud-

ies have investigated multimodal deep learning methods. These

models integrate data from diﬀerent neuroimaging modalities

(e.g., structural MRI, functional MRI, and PET scans) to create

a more comprehensive and robust diagnostic picture[16]. The

sheer volume of research has produced a rich body of literature

that systematically reviews the diverse array of AD detection

techniques, highlighting the consistent and rapid progress in di-

agnostic accuracy and model sophistication[17]. Furthermore,

the application of these powerful models has expanded beyond

AD to include the diagnosis of related neurodegenerative condi-

tions, such as tauopathies, thereby demonstrating their broader

versatility and clinical potential[18].

Since the ﬁeld has become mature, the quest to achieve

increasingly high accuracy has given rise to the emergence

of new and complicated model architectures. Hybrid systems

and ensemble-based systems are emerging to the fore. These

advanced approaches combine the predictive strengths of mul-

tiple deep learning models to create more robust, reliable, and

generalizable identiﬁcation systems that are less prone to the

biases of a single model[19, 20]. While neuroimaging remains

the primary data source, some innovative models have been de-

signed to diagnose AD based exclusively on structured patient

clinical records, providing a valuable alternative or comple-

mentary diagnostic pathway that does not require imaging[21].

This has contributed to the development of sophisticated ar-

chitectures capable not only of binary classiﬁcation but also of

detecting and diﬀerentiating between the various stages of AD,

from early-stage Mild Cognitive Impairment (MCI) to advanced

dementia, which is crucial for tailoring patient care[22, 23].

For these powerful deep learning models to transition from

research laboratories to real-world clinical environments, two

practical factors have become paramount: computational eﬃ-

ciency and model interpretability. Clinically viable diagnostic

tools have been developed to continue to focus on the trans-

fer learning application. The ﬁne-tuning of large models, which

have been trained on general image datasets, on smaller medical

imaging datasets is a very eﬀective way to achieve state-of-the-

art performance with small datasets[24]. This strategy has been

a central theme in the broader investigation of deep learning

for enhancing early detection and supporting clinical decision-

making[25, 26]. Consequently, a signiﬁcant portion of modern

research is focused on developing models for early detection, as

this is the stage at which interventions are most likely to be

eﬀective[27]. To further improve performance, some researchers

have proposed integrative models that combine the strengths of

traditional ML with advanced deep learning architectures[28].

Finally, to address the ”black-box” problem, there is a growing

emphasis on creating explainable AI (XAI). The development of

specialized, attention-based explainable networks and custom

models, such as ADD-Net, underscores the ﬁeld’s commit-

ment to creating diagnostic tools that are not only accurate

and eﬃcient but also transparent and trustworthy for clinical

use[29, 30].

Although many researchers have reported positive results

in detecting Alzheimer’s disease using deep learning models

in their literature, a wide range of methodological variations

can be seen in their approaches. Most of the existing liter-

ature has focused on MRI-based deep learning models using

CNN-based architectures and has achieved good accuracy us-

ing ADNI databases, ranging from 85% to 96%. However, a

few researchers have also used multimodal data for the ac-

curate prediction of Alzheimer’s disease using MRI and PET

scans. However, these techniques require expensive hardware.

In addition, fewer researchers have focused on CT-based deep

learning models for detecting Alzheimer’s disease. Moreover,

few researchers have focused only on prediction models without

using interpretability mechanisms or deployment frameworks.

Although various researchers have used various models, such as

ADD-Net and attention-based models, to explain their models,

these models are not used in deployment frameworks. However,

in the proposed model, a CT-based deep learning model was

used in conjunction with a structured clinical data model, an

explainable AI model using a Grad-CAM mechanism, and a

user-centric deployment model using a GUI and an AI chatbot

support system[31].

3. Methodology and Experiment

The proposed framework integrates a dual-model deep learning

pipeline for AD classiﬁcation with an interactive user interface

and an AI chatbot. The architecture is designed for accuracy,

interpretability, and user engagement.

3.1. Dataset and System Architecture

The general procedure of our system is illustrated in Fig-

ure 1. The training and evaluation dataset consists of 10,240

brain CT slices, evenly distributed between Alzheimer’s and

non-demented cases. In addition, a clinical dataset containing

2,149 patient records was used for structured data analysis.

Figure 2 presents the distribution of diagnoses in the clini-

cal dataset, with 64.6% classiﬁed as Non-Demented and 35.4%

as Demented. All the images were resized to an average of

224 × 224 pixels and normalised. Random rotations, horizontal

ﬂips, and zooming were used as data augmentation techniques

to improve the model’s robustness.

Figure 1. Alzheimer’s Detection System Architecture.

A stringent, automated preprocessing pipeline was applied

to the clinical data to prepare it for the FNN model. The ﬁrst

split of the features was into numerical (e.g., Age, MMSE) and

categorical ones (e.g., Gender, Smoking status). One-hot en-

coding was then applied to all categorical features in order to

put them in numerical form, and all numerical features were

rescaled to a normalized range. This ensures that all features

of the corresponding role in the prediction of the model are not

aﬀected by diﬀerent scales. All this transformation process was

stored as a single object of pipeline in order to make sure that

Shehu Mohammed et al.

Figure 2. Diagnosis Distribution in the Clinical Dataset.

whenever new information is keyed in the prediction process, it

would be processed similarly to the training information.

The CT imaging dataset used in this study consisted of ap-

proximately 30,000 brain CT slices derived from the Shrimaan

Super Specialty Hospital repository and clinical datasets for

Alzheimer’s disease research. Of these, 10,240 CT slices were

selected for training and evaluation purposes. The images were

divided into training, validation, and testing subsets using a

70–10–20 split. In addition, an extended testing pool was used

during the ﬁnal evaluation, resulting in approximately 4,500

CT slices used for performance assessment. Each CT slice in the

dataset was linked to a subject-level diagnosis label: Demented

or Non-Demented.

To avoid data leakage in the model, the clinical dataset was

split at the patient level instead of the slice level. This implies

that brain CT slices for a speciﬁc patient were only included

in either the training or testing dataset. Finally, the clinical

dataset was split into 70% for training, 10% for validation, and

20% for testing.

The clinical dataset included 2,149 patient records, which

included various physiological measures, cognitive measures

such as Mini-Mental State Examination (MMSE), and lifestyle

factors. For handling missing values in the clinical dataset, me-

dian imputation was applied for numerical features and most

frequent imputation for categorical features. This particular

process guaranteed the proper transformation of both training

and testing data.

3.2. System Deployment Architecture

The framework for deployment is modular, designed to function

in real-world environments. It has four components. The ﬁrst

is the user interface, the second is the inference engine, the

third is the data management system, and the fourth is the AI

chatbot module.

The primary interface with the system is the Graphical User

Interface (GUI). In the web-based system, the user can input

the CT scan of the brain or the clinical parameters. This is

then sent to the model inference server, which then uses the

EﬃcientNetV2-S and the FNN models to perform the predic-

tion. The EﬃcientNetV2-S is used to perform the prediction

with the CT scan, while the FNN is used with the clinical

parameters.

There is also an AI chatbot module that interacts with the

user in natural language to ask questions and educate the user

about Alzheimer’s disease. This module is known as the Neu-

roBot. The backend is designed to support an anonymized

clinical dataset and to perform inference requests with the

models.

The System deployment architecture of the proposed

Alzheimer’s detection framework is illustrated in Figure 3.

The system integrates a web-based user interface with back-

end APIs, deep learning inference models (EﬃcientNetV2-S for

brain CT slices and FNN for clinical data), and a chatbot mod-

ule to support clinical decision-making and patient interaction.

Figure 3. System Deployment Architecture Diagram.

3.3. Diagnostic Model Architectures

To create a comprehensive diagnostic tool, we employ two types

of neural nets: Convolutional Neural Networks (CNNs) to pro-

cess image data and Feedforward Neural Networks (FNNs) to

process data based on clinical features.

3.3.1. Architectures for Image-Based Analysis (CNNs)

The fundamental operation in a CNN is the two-dimensional

convolution. Convolution is a method of feature extraction, in

which a kernel is applied to a given input image or feature map.

This is arithmetically measured as:

O(i, j) = (I ∗ K)(i, j) =

I(i − m, j − n) · K(m, n) (1)

O(i, j) characterises the output feature map of location (i, j),

I is the input, and K is the convolutional kernel. Activation

functionality is applied to add non-linearity. Our application is

the ReLU, which is:

f(x) = max(0, x) (2)

For the image classiﬁcation task, this study utilizes the

EﬃcientNetV2-S architecture, a powerful and computationally

eﬃcient model. This model achieves its eﬃciency through the

use of Fused-MBConv blocks in its early layers.

Fused-MBConv block simpliﬁes the normal inverted resid-

ual block by fusing the ﬁrst 3 × 3 depthwise convolution of the

block with the second 1 × 1 projection convolution into a stan-

dard 3 × 3 convolution block. This minimizes the production

of memory access overhead increases and enhances the training

speed on advanced accelerators. The work of a Fused-MBConv

block could be explained by the equation:

y = BN (Conv

3×3

(BN (Conv

1×1

(x)))) + x (3)

In this case, x is a signal sent into the block, and the equa-

tion shows the series of a 1 × 1 expansion convolution, a 3 × 3

standard convolution (a replacement of the individual depth-

wise and projection steps) and Batch Normalization (BN). The

end result y is gotten by summing the original input x via a

residual (skip) connection, which makes this a distinguishing

feature that makes it likely to train very deep networks.

3.3.2. Architecture for Clinical Data Analysis (FNN)

A Feedforward Neural Network (FNN) was developed to ana-

lyze structured clinical data. This kind of network was trained

using a hyperparameter optimization strategy that entailed the

optimization of the predictive performance using the Optuna

framework.

The FNN is made up of a series of dense (fully connected)

layers. The general unit is its dense layer, which is a linear

transform of its input vector x, which can be expressed as:

y = W x + b (4)

The input is represented by y, the learnable weight matrix is

W , and the learnable bias is b.

The result of each of the dense layers is then run through

the activation function of the Rectiﬁed Linear Unit (ReLU) to

introduce non-linearity, as well as permit the model to learn

more intricate patterns, and is deﬁned as:

f(x) = max(0, x) (5)

Where γ and β are learnt scale and shift parameters, and the

constant ϵ is a small number to maintain numerical stability.

This is followed by the normalized output going through the

(ReLU) activation function.

To avoid overﬁtting, a ReLU activation will be followed by

a Dropout layer. Dropout is random and assigns a part of the

input units to 0 at every update time throughout the training

period, which assists in augmenting the strength of the network.

The optimized architecture that was discovered by hy-

perparameterOptimization is an input layer and two hidden

layers:

1. The initial hidden layer is a dense layer that contains 57

units, an activation of ReLU, and a Dropout with a rate

of 0.38.

2. The second hidden layer is a dense layer with 129 units,

and the next layer consists of a ReLU activation and a

Dropout layer at a rate of 0.49.

The concluding component is a single output neuron that

produces a raw logit, z. This logit is then transformed into

a probability using the sigmoid activation function. For im-

proved numerical stability during training, this function is

integrated directly into the loss function (BCEWithLogitsLoss).

The sigmoid function is deﬁned as:

σ(z) =

1 + e

−z

(6)

3.4. Interactive Framework and Training Algorithm

The project’s workﬂow, executed within a Jupyter Notebook,

provides two distinct diagnostic pathways: one for image-based

analysis and another for clinical feature-based analysis. The

classiﬁcation pipeline for the image-based pathway is detailed

in Algorithm 1, while the corresponding process for the clinical

data FNN model is outlined in Algorithm 2.

Algorithm 1 AD Classiﬁcation Pipeline (Image-Based CNN)

1: Input: Raw brain CT image I

raw

2: Data Loading: The image dataset is loaded from the

data/image directory and split into training and validation

sets.

3: Preprocessing & Augmentation:

4: I

resized

⇐ RandomResizedCrop(I

raw

, (224, 224))

5: I

flipped

⇐ RandomHorizontalFlip(I

resized

)

6: I

tensor

⇐ ToTensor(I

flipped

)

7: I

norm

⇐ Normalize(I

tensor

)

8: Model Loading: Load pre-trained model M

(EﬃcientNetV2-S) and adapt its ﬁnal layer for binary

classiﬁcation.

9: Training: The model is ﬁne-tuned on the training set using

an Adam optimizer and Binary Cross-Entropy with Logits

loss (BCEWithLogitsLoss).

10: Prediction:

11: Z ⇐ M(I

norm

) // Get raw logit output from the CNN

12: Classiﬁcation:

13: P ⇐ σ(Z) // Apply sigmoid function to get probability

14: If P > 0.5, then C ⇐ ’Demented’

15: Else C ⇐ ’Non-Demented’

16: Output: Return class label C.

Shehu Mohammed et al.

Algorithm 2 Clinical Data Classiﬁcation Pipeline (FNN)

1: Input: Raw clinical feature set D

raw

from

alzheimers disease data.csv.

2: Data Splitting: The dataset is split into training and testing

sets (80/20 split).

3: Preprocessing:

4: A ColumnTransformer pipeline is ﬁtted on the training

set:

5: Numerical features are imputed (median) and stan-

dardized (StandardScaler).

6: Categorical features are imputed (most frequent) and

one-hot encoded (OneHotEncoder).

7: D

train processed

⇐ Apply ﬁtted pipeline to training data.

8: D

test processed

⇐ Apply ﬁtted pipeline to test data.

9: Class Imbalance Handling:

10: D

train resampled

⇐ Apply SMOTE (Synthetic Minority

Over-sampling Technique) to D

train processed

to balance the

classes.

11: Model Loading: Load the pre-trained Feedforward Neural

Network (FNN).

12: Tensor Conversion:

13: T

input

⇐ ToTensor(D

test processed

)

14: Prediction:

15: Z ⇐ M

fnn

input

) // Get raw logit output from the FNN

16: Classiﬁcation:

17: P ⇐ σ(Z) // Apply sigmoid function to get probability

18: If P > 0.5, then C ⇐ ’Demented’

19: Else C ⇐ ’Non-Demented’

20: Output: Return class label C.

4. Results and Discussion

The eﬃcacy of the hybrid framework that was suggested was

critically evaluated using the aid of an integrated approach.

This was accompanied by a quantitative measure of the classiﬁ-

cation performance of the deep learning models using a held-out

test set of 4,500 brain CT slices, a large-scale measure of the

associated clinical data to validate the validity of the features of

the models, as well as a qualitative measure of the user interface

and the interactivity aspects.

4.1. Performance Metrics

We used quantitative measures of the models using a sam-

ple set of standard classiﬁcation measures that is decided by

the elements of the confusion matrix: True Positives (TP),

True Negatives (TN), False Positives (FP), and False Negatives

(FN).

• Accuracy: The percentage of all the correct predictions.

Accuracy =

T P + T N

T P + T N + F P + F N

(7)

• Precision: The proportion of positive predictions that were

actually correct.

P recision =

T P

T P + F P

(8)

• Recall (Sensitivity): The proportion of actual positives

that were identiﬁed correctly.

Recall =

T P

T P + F N

(9)

• F1-Score: The harmonic mean of Precision and Recall,

providing a single score that balances both.

F 1-Score =

2 · P recision · Recall

P recision + Recall

(10)

4.2. Model Interpretability and Performance

Evaluation

Besides the model’s accuracy, interpretability is another critical

requirement for clinical decision-support system. In this study,

the interpretability of the model was facilitated by the use of a

technique called Gradient-weighted Class Activation Mapping

(Grad-CAM) in the EﬃcientNetV2-S CNN model. This tech-

nique enables the production of visual maps that highlight areas

of the images used in the model that are being used to make a

prediction, as illustrated in Figure 4. These maps allow veriﬁ-

cation that the model focuses on anatomically relevant regions

associated with Alzheimer’s disease, such as cortical atrophy

and ventricular enlargement.

Although Grad-CAM visualizations, statistical validation,

and feature-importance analysis provide valuable insights into

the model’s decision-making process. Future work will investi-

gate quantitative explainability metrics, including localization-

based evaluation and clinician-assisted validation of explana-

tion maps, to provide a more rigorous assessment of model

interpretability and clinical trustworthiness.

Moreover, the interpretability of the clinical data pathway

was facilitated by the use of statistical tests, such as indepen-

dent sample t-tests and correlation tests, which showed that

clinical variables, such as MMSE, Activities of Daily Living

(ADL), and functional assessment, signiﬁcantly vary across the

diﬀerent diagnostic groups.

Figure 4. Grad-CAM visualization highlighting important regions in

brain CT slices used by the EﬃcientNetV2-S model for Alzheimer’s

disease classiﬁcation. The highlighted areas indicate regions contributing

most strongly to the model’s prediction.

The EﬃcientNetV2-S model achieved the best validation

performance with an accuracy of 98.19% compared to other

architectures, compared with other evaluated architectures,

including EﬃcientNetV2-B0, ResNet50V2, and VGG16, as in-

dicated in Table 1. Figure 5 shows the visualization of the

performance of these models in comparison with each other, and

it is evident that the EﬃcientNetV2-S model is more accurate.

Figure 6 displays the training and validation history of the

EﬃcientNetV2-S model and depicts the change in the accu-

racy and loss after 10 epochs. This trend is convergently stable,

and the trend has small overﬁtting, which is an attribute that

indicates good ﬁne-tuning and regularization.

Model Accuracy (%)

EﬃcientNetV2-S 98.19%

EﬃcientNetV2-B0 95.8%

ResNet50V2 94.9%

VGG16 91.2%

DenseNet121 90.7%

MobileNetV2 89.3%

Table 1. Performance Comparison of Deep Learning Models

Figure 5. Visualization of Performance Comparison of All Models.

Figure 6. Training and Validation History for the EﬃcientNetV2-S

Model, showing Loss and Accuracy over 10 epochs.

As a method of measuring the reliability of classiﬁcations,

a confusion matrix of the EﬃcientNetV2-S model was con-

structed and presented in Figure 7. The matrix shows the

accuracy of the model and recall of both the Demented and

Non-Demented classes, that provide us with insight into the

discriminatory ability of the model.

The FNN model, after undergoing hyperparameter tuning

with Optuna, achieved a ﬁnal test accuracy of 87.21% on the

clinical dataset. This result validates the eﬀectiveness of using

structured clinical data for prediction and highlights the beneﬁt

of automated hyperparameter optimization.

In the dual-modality framework, there is the incorporation

of medical images and clinical data, which makes the diagno-

sis even better. The CNN examines the images, digging deeper

into the changes seen on the CT scan, such as cortical atrophy

and enlarged ventricles, which indicate neurodegeneration. On

the other hand, the FNN examines the clinical data, which in-

cludes cognitive tests such as MMSE, daily living activities, and

other relevant demographic factors that may be associated with

the condition. Statistical analysis reveals that MMSE, ADL,

and other functional measures vary signiﬁcantly across diﬀer-

ent diagnostic groups. The dual-modality framework, therefore,

examines both visible changes on the brain images and the pa-

tient’s individual risk factors, making the diagnosis even better

and more reliable.

Figure 7. Confusion Matrix for EﬃcientNetV2-S on the Test Set.

To further evaluate the eﬀectiveness of the proposed frame-

work, its performance was compared with several recent state-

of-the-art approaches for Alzheimer’s detection reported in the

literature. These studies employed various deep learning archi-

tectures and imaging modalities, including MRI-based convolu-

tional neural networks and multimodal models. The comparison

result presented in Table 2 demonstrates that the proposed

framework achieves competitive performance while integrating

CT imaging, structured clinical data, and an interactive clinical

support system.

It should be noted that direct comparison of classiﬁcation

accuracies across studies should be interpreted with caution be-

cause the evaluated datasets, imaging modalities, sample sizes,

and experimental protocols diﬀer substantially. Most of the

compared studies employed MRI-based datasets, particularly

ADNI, whereas the proposed framework was developed using

brain CT images and structured clinical data obtained from a

hospital-based cohort. MRI generally provides higher soft-tissue

contrast than CT imaging, which may inﬂuence diagnostic

performance, Furthermore, variations in dataset composition,

preprocessing procedures, and evaluation strategies can aﬀect

reported accuracies. Despite these diﬀerences, the proposed

framework achieved competitive performance while simulta-

neously providing explainability through Grad-CAM and a

deployment-oriented clinical support platform.

4.3. User Interface and System Interaction

The practical utility of the diagnostic models is realized through

an intuitive and user-centric graphical user interface (GUI). The

system’s main dashboard, shown in Figure 8, provides a clean

and accessible entry point for users. It presents two primary

options for analysis, feature-based (clinical data) and image-

based, alongside the integrated ”NeuroBot” AI assistant.

For a feature-based analysis, the user navigates to a com-

prehensive data input form (Table 3), where they can enter

demographic and clinical parameters. After submission, the sys-

tem presents a detailed results page (Figure 9) that not only

displays the diagnosis (’Demented’ or ’Non-Demented’) but also

provides actionable suggestions and precautions tailored to the

outcome. This aspect turns the system into a supportive system

rather than a basic classiﬁer. For an image-based diagnosis, the

Shehu Mohammed et al.

Study Data Modality Model / Method Dataset Accuracy (%)

Pradhan et al., 2024 MRI DenseNet + ResNet50 ADNI 95.3

Helaly et al., 2022 MRI CNN-based Deep Learning ADNI 93.6

Saleem et al., 2022 MRI Transfer Learning CNN ADNI 94.7

Ayus & Gupta, 2024 MRI Hybrid Ensemble DL ADNI 96.2

Alwakid et al., 2024 MRI Image Processing + CNN ADNI 94.5

Avila-Jim´enez et al., 2024 Clinical Records Deep Learning Model Clinical Dataset 85.0

Proposed Method CT + Clinical Data EﬃcientNetV2-S + FNN Hospital + Clinical Dataset 98.19

Table 2. Comparison of the proposed method with recent state-of-the-art Alzheimer’s disease detection approaches.

Figure 8. Main Dashboard Interface with Chatbot Panel.

user interacts with a simple drag-and-drop interface to upload

a brain CT scan (Figure 10). The system processes the image

and returns a clear results page displaying the diagnosis, a con-

ﬁdence score, and preventive measures or daily activities that

may be beneﬁcial (Figure 11). This consistent, informative clin-

ical workﬂow is formulated to be easily ventured by clinicians,

patients, and even caregivers, and this gap between a sophis-

ticated AI model and a practical and usable tool is addressed.

4.4. Statistical Validation and Feature Analysis

To oﬀer a sound clinical basis to our models, the struc-

tured clinical dataset underwent a complete statistical analysis.

This validation guarantees that the characteristics learnt by

the models are associated with accepted clinical markers of

Alzheimer’s disease.

This was done by the use of an independent samples t-test

to identify the clinical characteristics that signiﬁcantly diﬀered

between the Demented and Non-Demented groups. Figure 12

gives the general ﬁndings of the t-test in a manner that sum-

marizes the comparative statistics of the two diagnostic groups.

In the analysis, it was found that the signiﬁcant indicators that

constitute an overwhelming proportion of the signiﬁcant indi-

cators were the MMSE score, and that the p-value was 0.0000.

These diﬀerences in distribution between the Age and MMSE

age scores, as represented graphically in Figure 13 and Fig-

ure 14 of the boxplot is a very sharp and signiﬁcant drop in

MMSE scores in the demented group compared with Age.

The correlation heatmap in Figure 15 also shows correlations

between major clinical characteristics and the ﬁnal diagnosis,

indicating that the MMSE score is most strongly negatively

correlated with dementia diagnosis.

Most importantly, all these statistics are directly related to

the feature-importance analysis of the FNN model presented

in Figure 16. These characteristics, the FunctionalAssessment,

ADL (Activities of Daily Living), MemoryComplaints and

MMSE, were evaluated as the most eﬀective predictors for the

model. Such a high concordance rate assures that the FNN

model evidently learns to focus on the clinically signiﬁcant

variables associated with Alzheimer’s disease.

In addition to the visual explanations provided by Grad-

CAM, quantitative evidence supporting model interpretabil-

ity was obtained through statistical validation of the clinical

variables. Independent sample t-tests demonstrated signiﬁ-

cant diﬀerences between demented and Non-demented groups,

with MMSE exhibiting a highly signiﬁcant association (p <

0.001). Correlation analysis further conﬁrmed strong rela-

tionships between important clinical features and dementia

Feature Value Feature Value

Age 68 SystolicBP 118

Gender 0 DiastolicBP 78

Ethnicity 0 CholesterolTotal 180

EducationLevel 3 CholesterolLDL 100

BMI 24.5 CholesterolHDL 65

Smoking 0 CholesterolTriglycerides 130

AlcoholConsumption 2 MMSE 29

PhysicalActivity 56 FunctionalAssessment 1.0

DietQuality 9 MemoryComplaints 0

SleepQuality 8 BehavioralProblems 0

FamilyHistoryAlzheimers 0 ADL 1.0

CardiovascularDisease 0 Confusion 0

Diabetes 0 Disorientation 0

Depression 0 Personality Changes 0

HeadInjury 0 Diﬃculty Completing Tasks 0

Hypertension 0 Forgetfulness 1

Doctor In Charge 221

Table 3. Sample Input Values for Feature-Based Analysis.

Figure 9. Analysis Results Page for Clinical Data Submission.

diagnosis. Furthermore, feature-importance analysis identiﬁed

MMSE, Functional Assessment, ADL, and Memory Complaints

as the most inﬂuential used by the FNN model. The agreement

between statistical signiﬁcance and model-derived feature im-

portance provides quantitative evidence that the model focuses

on clinically meaningful biomarkers associated with Alzheimer’s

disease.

4.5. User Engagement Framework

The ﬁnal component of our evaluation focused on the AI-

powered chatbot, NeuroBot, which serves as the primary tool

for patient and caregiver support. The chatbot was assessed

across ﬁve key criteria: Domain Relevance, Medical Accuracy,

Politeness and Safety, Responsiveness, and its ability to refuse

out-of-scope questions. As illustrated in the radar chart in

Figure 17, NeuroBot scored perfectly or near-perfectly on all

measures. This lends credibility to its accuracy, safety, and rel-

evancy in delivering pertinent information within the ﬁeld of

Alzheimer Disease. Also, response time analysis revealed that

the response to user queries was received in time (85% of queries

were attended to within a range of 4.2 to 4.6 seconds), which

guaranteed a consistent and interactive user experience.

5. Conclusion and Future Work

This study developed and evaluated a dual-modality framework

for the early detection of Alzheimer’s disease, where the deep

learning pipeline is applied to the CT image processing and the

hyperparameter-optimized Feedforward Neural Network (FNN)

is applied to the clinical data processing.

The ﬁndings demonstrate that using a ﬁne-tuned Eﬃ-

cientNetV2-S architecture for image analysis can achieve high

diagnostic accuracy, reaching a peak validation performance

of 98.19%. Complementing this, the FNN model, optimized

through systematic hyperparameter tuning, achieved a ﬁnal

accuracy of 87.21% on structured clinical data. The statisti-

cal analysis of this clinical data further solidiﬁed our approach,

conﬁrming that the models are learning from features, such as

the MMSE score, that are strongly correlated with established

indicators of cognitive decline.

One of the major contributions that this research makes is

the holistic approach. When putting these diagnostic models

into a broader ecosystem, which includes a user-friendly inter-

face and the AI-powered NeuroBot, this piece of work oﬀers a

blueprint of an all-inclusive support platform. This usability-

based strategy is also essential in translating the gap between

the complicated AI technology and the daily clinical practice,

Shehu Mohammed et al.

Figure 10. Drag-and-Drop Interface for Image-Based Analysis.

Figure 11. Analysis Results Page for Image-Based Diagnosis.

and allows patients, caregivers, and clinicians to place the right

diagnostics and information right at their ﬁngertips.

The proposed framework oﬀers promising avenues for future

development and presents several opportunities for improve-

ment. The capabilities of the diagnostic functions will be

expanded to provide a more comprehensive portrait of the pa-

tient, with a priority on the integration of multimodal data from

MRI scans and genetic markers. We will also make the existing

binary classiﬁcation models extendable so that Alzheimer’s can

be classiﬁed in multiple steps, so that we are in a position to

further distinguish between MCI and the onset of further stages

of the disease.

The last and most important stage will be the push of the

framework into clinical validation by means of large-scale trials.

This will be essential for validating the system’s performance

across diverse populations and is a necessary step for its even-

tual integration into standard clinical workﬂows. Concurrently,

we will explore the optimization of the models for deployment

on edge devices and expand the chatbot’s capabilities to include

summarizing prediction results and oﬀering multilingual sup-

port, thereby increasing the system’s accessibility and global

impact.

It should be noted that the above evaluation was carried out

using controlled experimental data. Although the models were

Figure 12. Visualization of T-Test Results Comparing Demented vs. Non-Demented Groups.

Figure 13. Boxplot of Age Distribution by Diagnosis.

validated using a held-out test set, the models were not vali-

dated using any independent clinical data sets. In the future,

the focus will be on evaluating the framework using data sets

from multiple institutions and clinical settings.

Shehu Mohammed et al.

Figure 14. MMSE Score Distribution by Diagnosis (T-Test Validation).

Figure 15. Correlation Heatmap of Top Clinical Features

Figure 16. Top 10 Most Important Features.

Figure 17. Radar Chart Evaluation of the Chatbot.

Shehu Mohammed et al.

Ethical Statement

Ethical approval of this study was obtained from the Insti-

tutional Ethics Committee of Lovely Professional University,

India (Ref: LPU/IEC-LPU/2025/1/2, dated 15 February 2025).

Permission to access clinical and imaging data was granted

by Shrimaan Superspeciality Hospital, Jalandhar, India. The

study was retrospective and involved analysis of previously col-

lected anonymised clinical and brain CT imaging data. No

direct patient contact occurred, and no personally identiﬁable

information was accessed. In accordance with the Institutional

Ethics Committee clariﬁcation, the requirement for informed

consent was waived/not applicable.

This study employed anonymized brain images from the

brain CT slices of patients at Shrimann Superspeciality Hospi-

tal. No personally identiﬁable patient information was accessed

during this study. The data were used strictly for academic

research in accordance with the hospital’s data privacy regula-

tions. The proposed framework is intended for use as a clinical

decision support tool. However, it is not intended for use as

a substitute for clinical expertise. The NeuroBot chatbot pro-

vides general information guidance only and does not replace

professional advice.

Funding

The scholar was sponsored by Tertiary Education Trust Fund

(TETFUND) Nigeria for Higher education Studies only.

Declaration of competing interests

The authors declare that they have no known competing ﬁnan-

cial interests or personal relationships that could have appeared

to inﬂuence the work reported in this paper.

Data Availability Statements

The data supporting the ﬁndings of this study are available

from the corresponding author upon reasonable request. How-

ever, the data are not publicly available due to privacy or

ethical restrictions.

Credit authorship contribution statement

Shehu Mohammed: Conceptualization; Project Administration;

Methodology; Data Curation; Software Development; Investi-

gation; Writing – Original Draft Preparation.

Neha Malhotra: Supervision; Writing – Review & Editing; For-

mal Analysis.

Anmol Singh Rai: Resources; Validation; Visualization.

References

1. S. Alinsaif, “Dca-enhanced alzheimer’s detection with

shearlet and deep learning integration,” Computers in

Biology and Medicine, vol. 185, p. 109538, 2025, doi:

https://doi.org/10.1016/j.compbiomed.2024.109538.

2. M. Srikanth and P. Bellapukonda, “The early detection of

alzheimer’s illness using machine learning and deep learning

algorithms,” Journal of Pharmaceutical Negative Results,

vol. 13, no. 9, pp. 4852–4859, 2022, doi: https://doi.org/

10.47750/pnr.2022.13.S09.603. [Online]. Available: https://

www.pnrjournal.com/index.php/home/article/view/4470

3. E. M. Mohammed, A. M. Fakhrudeen, and O. Y. Alani,

“Detection of alzheimer’s disease using deep learning

models: A systematic literature review,” Informatics

in Medicine Unlocked, vol. 50, p. 101551, 2024, doi:

https://doi.org/10.1016/j.imu.2024.101551.

4. A. D. Arya et al., “A systematic review on machine

learning and deep learning techniques in the eﬀective

diagnosis of alzheimer’s disease,” Brain Informatics,

vol. 10, no. 1, p. 17, 2023, doi: https://doi.org/10.1186/

s40708-023-00195-7.

5. N. Pradhan, S. Sagar, and A. S. Singh, “Analysis

of mri image data for alzheimer disease detection

using deep learning techniques,” Multimedia Tools and

Applications, vol. 83, no. 6, pp. 17 729–17 752, 2024, doi:

https://doi.org/10.1007/s11042-023-16256-2.

6. P. Kishore, C. U. Kumari, M. Kumar, and T. Pavani,

“Detection and analysis of alzheimer’s disease using

various machine learning algorithms,” Materials today:

proceedings, vol. 45, pp. 1502–1508, 2021, doi: https:

//doi.org/10.1016/j.matpr.2020.07.645.

7. P. Balaji, M. A. Chaurasia, S. M. Bilfaqih, A. Muniasamy,

and L. E. G. Alsid, “Hybridized deep learning approach

for detecting alzheimer’s disease,” Biomedicines, vol. 11,

no. 1, p. 149, 2023, doi: https://doi.org/10.3390/

biomedicines11010149.

8. S. Mirabian, F. Mohammadian, Z. Ganji, H. Zare, and

E. Hasanpour Khalesi, “The potential role of machine

learning and deep learning in diﬀerential diagnosis of

alzheimer’s disease and ftd using imaging biomarkers:

A review,” The Neuroradiology Journal, vol. 38,

no. 5, pp. 571–587, 2025, doi: https://doi.org/10.1177/

19714009251313511.

9. J. Chua et al., “Utilizing deep learning to predict

alzheimer’s disease and mild cognitive impairment with

optical coherence tomography,” Alzheimer’s & Dementia:

Diagnosis, Assessment & Disease Monitoring, vol. 17,

no. 1, p. e70041, 2025, doi: https://doi.org/10.1002/dad2.

70041.

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

and M. Hammad, “Accurate detection of alzheimer’s

disease using lightweight deep learning model on mri

data,” Diagnostics, vol. 13, no. 7, p. 1216, 2023, doi:

https://doi.org/10.3390/diagnostics13071216.

11. K. Lokesh, N. P. Challa, A. S. Satwik, J. C. Kiran, N. K.

Rao, and B. Naseeba, “Early alzheimer’s disease detection

using deep learning,” EAI Endorsed Transactions on

Pervasive Health and Technology, vol. 9, 2023, doi:

https://doi.org/10.4108/eetpht.9.3966.

12. S. Balne and A. Elumalai, “Machine learning and deep

learning algorithms used to diagnosis of alzheimer’s,”

Materials Today: Proceedings, vol. 47, pp. 5151–5156,

2021, doi: https://doi.org/10.1016/j.matpr.2021.05.499.

13. S. Baskar, M. L. Prasad, N. Sharma, T. Katale, P. C. S.

Reddy et al., “An accurate prediction and diagnosis

of alzheimer’s disease using deep learning,” in 2023

IEEE North Karnataka Subsection Flagship International

Conference (NKCon). IEEE, 2023, pp. 1–7, doi:

https://doi.org/10.1109/nkcon59507.2023.10396132.

14. A. Muydinov, “Advances in deep learning techniques for

alzheimer’s disease detection using mri images review,” in

Proceedings of the 8th International Conference on Future

Networks & Distributed Systems, 2024, pp. 266–269, doi:

https://doi.org/10.1145/3726122.3726160.

15. G. N. Alwakid, S. Tahir, M. Humayun, and W. Gouda,

“Improving alzheimer’s detection with deep learning and

image processing techniques,” IEEE Access, vol. 12,

pp. 153 445–153 456, 2024, doi: https://doi.org/10.1109/

ACCESS.2024.3481238.

16. G. S. Chhabra, A. Guru, B. J. Rajput, L. Dewangan,

and S. K. Swarnkar, “Multimodal neuroimaging for

early alzheimer’s detection: a deep learning approach,”

in 2023 14th international conference on computing

communication and networking technologies (ICCCNT).

IEEE, 2023, pp. 1–5, doi: https://doi.org/10.1109/

ICCCNT56998.2023.10307780.

17. S. Afzal et al., “Alzheimer disease detection techniques and

methods: A review.” International journal of interactive

multimedia and artiﬁcial intelligence, vol. 6, no. 7, pp.

26–38, 2021, doi: https://doi.org/10.9781/ijimai.2021.04.

005.

18. S. Koga, A. Ikeda, and D. W. Dickson, “Deep learning-

based model for diagnosing alzheimer’s disease and

tauopathies,” Neuropathology and Applied Neurobiology,

vol. 48, no. 1, p. e12759, 2022, doi: https://doi.org/10.

1111/nan.12759.

19. I. Ayus and D. Gupta, “A novel hybrid ensemble

based alzheimer’s identiﬁcation system using deep learning

technique,” Biomedical Signal Processing and Control,

vol. 92, p. 106079, 2024, doi: https://doi.org/10.1016/j.

bspc.2024.106079.

20. S. Suganyadevi, A. S. Rajasekaran, N. Satheesh, R. Sug-

anthi, R. Naveenkumar et al., “Alzheimer’s disease

diagnosis using deep learning approach,” in 2023 Second

International Conference on Electronics and Renewable

Systems (ICEARS). IEEE, 2023, pp. 1205–1209, doi:

https://doi.org/10.1109/ICEARS56392.2023.10085017.

21. J. L.

Avila-Jim´enez, V. Cant´on-Habas, M. del Pilar

Carrera-Gonz´alez, M. Rich-Ruiz, and S. Ventura, “A

deep learning model for alzheimer’s disease diagnosis

based on patient clinical records,” Computers in Biology

and Medicine, vol. 169, p. 107814, 2024, doi:

https://doi.org/10.1016/j.compbiomed.2023.107814.

22. S. Sava¸s, “Detecting the stages of alzheimer’s disease with

pre-trained deep learning architectures,” Arabian Journal

for Science and Engineering, vol. 47, no. 2, pp. 2201–2218,

2022, doi: https://doi.org/10.1007/s13369-021-06131-3.

23. W. Al Shehri, “Alzheimer’s disease diagnosis and

classiﬁcation using deep learning techniques,” PeerJ

Computer Science, vol. 8, p. e1177, 2022, doi:

https://doi.org/10.7717/peerj-cs.1177.

24. T. J. Saleem et al., “Deep learning-based diagnosis of

alzheimer’s disease,” Journal of Personalized Medicine,

vol. 12, no. 5, p. 815, 2022, doi: https://doi.org/10.3390/

jpm12050815.

25. G. Hcini, I. Jdey, and H. Dhahri, “Investigating

deep learning for early detection and decision-making in

alzheimer’s disease: A comprehensive review: G. hcini et

al.” Neural Processing Letters, vol. 56, no. 3, p. 153, 2024,

doi: https://doi.org/10.1007/s11063-024-11600-5.

26. S. Jain and R. Srivastava, “Enhanced eeg-based alzheimer’s

disease detection using synchrosqueezing transform and

deep transfer learning,” Neuroscience, vol. 576, pp.

105–117, 2025, doi: https://doi.org/10.1016/j.neuroscience.

2025.04.041.

27. H. A. Helaly, M. Badawy, and A. Y. Haikal, “Deep learning

approach for early detection of alzheimer’s disease,”

Cognitive computation, vol. 14, no. 5, pp. 1711–1727,

2022, doi: https://doi.org/10.1007/s12559-021-09946-2.

28. T. Vanaja, K. Shanmugavadivel, M. Subramanian, and

C. Kanimozhiselvi, “Advancing alzheimer’s detection:

integrative approaches in mri analysis with traditional

and deep learning models,” Neural Computing and

Applications, vol. 37, no. 14, pp. 8527–8546, 2025, doi:

https://doi.org/10.1007/s00521-025-10993-1.

29. M. M. S. Fareed et al., “Add-net: an eﬀective deep

learning model for early detection of alzheimer disease in

mri scans,” IEEE Access, vol. 10, pp. 96 930–96 951, 2022,

doi: https://doi.org/10.1109/ACCESS.2022.3204395.

30. E. Kina, “Tleablcnn: Brain and alzheimer’s disease

detection using attention based explainable deep learning

and smote using imbalanced brain mri,” IEEE Access,

2025, doi: https://doi.org/10.1109/ACCESS.2025.3539550.

31. M. Taiyeb Khosroshahi et al., “Explainable artiﬁcial

intelligence in neuroimaging of alzheimer’s disease,”

Diagnostics, vol. 15, no. 5, p. 612, 2025, doi:

https://doi.org/10.3390/diagnostics15050612.