Alzheimer Detection Via 2D Hybrid Neural Networks

A PROJECT REPORT

Submitted by

Vivek Chauhan - 22BCS15816

Aman Kumar– 22BCS15813
Sanchit Sharma – 22BCS15821
Yashas Bhardwaz – 22BCS15817

in partial fulfillment for the award of the degree of

BACHELOR OF ENGINEERING

IN

COMPUTER SCIENCE AND ENGINEERING

Chandigarh University
DEC-2025

1
 BONAFIDE CERTIFICATE

Certified that this project report "Alzheimer Detection Via 2D Hybrid Neural
Networks" is the bonafide work of “Vivek Chauhan - 22BCS15816,
Aman Kumar- 22BCS15813, Sanchit Sharma - 22BCS15821,
Yashas - 22BCS15817” who carried out the project work under my/our
supervision.

SIGNATURE SIGNATURE

HEAD OF THE DEPARTMENT SUPERVISOR

Prof(Dr.) Navpreet Kaur Walia Er. Shailja Saini

Submitted for the project viva-voice

Examination held on_

INTERNAL EXAMINER EXTERNAL EXAMINER

2
 TABLE OF CONTENTS

ABSTRACT …………………………………………………………..5
CHAPTER 1. INTRODUCTION
1.1 Overview of Alzheimer’s Disease ...................................................10
1.2 Need for AI-Powered Detection Tools ............................................................15
1.3 Role of Machine Learning and NLP in Fake News Detection ......................23
1.4 Objectives and Scope of the Study ..................................................................25
CHAPTER 2. LITERATURE REVIEW
2.1 Evolution of Fake News and Its Impact on Society .......................................27
2.2 Existing Fake News Detection Approaches ....................................................30
2.3 Gaps in Current Research ...............................................................................31
2.4 Relevance of AI in Combating Misinformation .............................................44
CHAPTER 3. METHODOLOGY
3.1 Data Collection and Preprocessing ..................................................................48
3.2 Feature Extraction using NLP Techniques .....................................................49
3.3 Machine Learning Algorithms for Classification ...........................................50
3.4 Evaluation Metrics for Fake News Detection ..................................................51
CHAPTER 4. IMPLEMENTATION
4.1 Tools and Technologies (Python, Scikit-learn, NLTK, TensorFlow) ...........53
4.2 System Architecture and Workflow ................................................................54
4.3 Model Training and Testing Process ...............................................................55
4.4 Example Case Study: WhatsApp/Facebook Headline Detection ..................56
CHAPTER 5. CONCLUSION AND FUTURE WORK
5.1 Summary of Findings ........................................................................................59
5.2 Key Contributions of the Study ........................................................................60
5.3 Recommendations for Future Work ................................................................62

REFERENCES ......................................................................................49

3
 ABSTRACT

The global rise in neurodegenerative disorders such as Alzheimer’s disease has underscored the
urgent need for intelligent and automated diagnostic tools capable of supporting early detection
and clinical decision-making. Alzheimer’s disease (AD) is a progressive neurological disorder that
gradually destroys brain cells, leading to cognitive impairment, memory loss, and eventual loss of
independent function. Traditional diagnostic techniques, primarily relying on neuropsychological
assessments and manual evaluation of magnetic resonance imaging (MRI) scans, often suffer from
subjectivity, time constraints, and limited scalability in clinical practice. In this context, the
integration of Artificial Intelligence (AI) and Deep Learning (DL) has emerged as a
transformative solution, offering the potential to detect subtle structural and textural abnormalities
in brain scans with remarkable precision and speed.

This research presents an advanced computational framework for Alzheimer Detection via 2D
Hybrid Neural Networks, designed to efficiently classify brain MRI images into distinct
categories representing different stages of the disease—Non-Demented, Very Mild Demented,
Mild Demented, and Moderate Demented. The proposed hybrid model synergistically combines
Convolutional Neural Networks (CNNs) for automated feature extraction with Fully Connected
Deep Neural Networks (DNNs) for robust classification. This hybridization leverages the spatial
feature learning capacity of CNNs while maintaining the generalization and interpretability
advantages of dense layers. Unlike conventional machine learning methods dependent on
handcrafted features, the system autonomously learns discriminative features from raw 2D MRI
inputs, thus minimizing human bias and improving diagnostic consistency.

The methodological pipeline encompasses multiple stages, ensuring data integrity and optimal
model performance. MRI scans were collected from publicly available repositories, including the
Alzheimer’s Disease Neuroimaging Initiative (ADNI) and the Kaggle Alzheimer MRI dataset.
Preprocessing techniques such as grayscale normalization, noise removal, histogram equalization,
resizing, and contrast enhancement were applied to standardize input quality. Furthermore, data
augmentation (rotation, flipping, and zooming) was employed to mitigate.

4
The model architecture was implemented using TensorFlow and Keras frameworks. It consists
of multiple convolutional blocks with ReLU activation functions and max-pooling layers to extract
low- and high-level spatial features from MRI slices. These are followed by flattening, dropout
regularization, and dense layers for classification. The final output layer uses the Softmax
activation function to produce probabilistic predictions across four Alzheimer stages. The Adam
optimizer with adaptive learning rates was used to accelerate convergence, while categorical
cross-entropy served as the loss function for multi-class classification.

The experimental design included a comprehensive evaluation of multiple architectures and

hyperparameters to determine the optimal configuration. Baseline CNN models were first trained
to establish comparative performance metrics. Subsequently, hybrid models incorporating
additional dense layers and regularization techniques demonstrated superior performance. The
dataset was divided into training (70%), validation (20%), and testing (10%) partitions to ensure
fair performance assessment and prevent data leakage.

Performance was measured using multiple statistical metrics—Accuracy, Precision, Recall, F1-
Score, and Confusion Matrix—to provide a holistic evaluation. The proposed 2D Hybrid Neural
Network achieved an overall classification accuracy of 93–95%, outperforming traditional CNN
and classical ML classifiers such as Random Forest and Support Vector Machines (SVM).
Moreover, the model achieved an F1-score exceeding 0.93, indicating balanced precision and
recall across all classes.

Beyond quantitative results, qualitative visualization techniques such as activation mapping and
heatmap overlays were employed to interpret the model’s decision-making process, highlighting
brain regions most influential in classification. This interpretability aspect is crucial for fostering
clinical trust and enabling AI-assisted decision support in medical contexts.

The findings of this study demonstrate that hybrid deep learning architectures can effectively
model the complex structural variations in MRI images associated with Alzheimer’s progression.
The hybrid approach optimizes the trade-off between computational efficiency and diagnostic
accuracy, making it suitable for both clinical research and real-time deployment scenarios.

5
From a societal perspective, the implications are significant. Early and accurate Alzheimer
detection through AI-driven tools can substantially improve patient outcomes by enabling timely
interventions and treatment strategies. In regions with limited access to medical specialists—
particularly in developing nations—such systems can function as cost-effective diagnostic aids,
bridging the gap between resource availability and healthcare needs.

However, several limitations are acknowledged. Model performance is sensitive to dataset

diversity, image quality, and annotation accuracy. The current framework is restricted to 2D MRI
scans, which may not fully capture volumetric brain patterns present in 3D imaging. Additionally,
large-scale clinical validation remains essential before real-world deployment.

Future research directions include extending the model to 3D CNN architectures for volumetric
analysis, integrating transfer learning with pre-trained models such as ResNet50, and
implementing Explainable AI (XAI) methods to enhance transparency and interpretability.
Further, incorporating multimodal data (e.g., PET scans, genetic information, cognitive scores)
and developing lightweight architectures optimized for edge computing and IoT-based
healthcare devices could significantly broaden the system’s applicability.

In conclusion, this study establishes a comprehensive, data-driven framework for automated

Alzheimer detection using 2D Hybrid Neural Networks. By uniting the strengths of convolutional
and dense architectures, the system delivers high accuracy, interpretability, and scalability. It not
only contributes to the ongoing evolution of medical image analysis but also demonstrates the
transformative potential of AI in addressing critical challenges in neurological healthcare—paving
the way for earlier diagnosis, improved patient care, and a more intelligent healthcare ecosystem.

6
 CHAPTER 1.
INTRODUCTION

1.1 Overview of Alzheimer’s Disease

Alzheimer’s disease (AD) represents one of the most profound medical, social, and economic
challenges of the 21st century. As a progressive neurodegenerative disorder characterized by
cognitive decline, memory impairment, and behavioral disturbances, Alzheimer’s disrupts the very
essence of human identity and independence. The disease gradually deteriorates neurons and
synapses in brain regions associated with memory and reasoning, ultimately leading to complete
dependency and death. With global populations aging at unprecedented rates, the incidence of
Alzheimer’s is rising rapidly—transforming from a medical condition into a societal crisis
demanding technological, medical, and ethical responses.

According to the World Health Organization (WHO, 2023), more than 55 million people
worldwide live with dementia, with Alzheimer’s disease accounting for nearly 60–70% of those
cases. By 2050, this number is projected to triple due to increased life expectancy and aging
populations, particularly in developing countries. Beyond the individual suffering, Alzheimer’s
imposes enormous economic burdens—both direct (healthcare costs, hospitalization) and indirect
(loss of caregiver productivity, long-term care facilities). The Alzheimer’s Association estimates
global economic losses exceeding $1.3 trillion annually, a figure expected to double by 2030.
These staggering statistics underline the urgency for early detection, which can slow disease
progression through medical intervention, behavioral therapies, and improved quality of life.

The term “Alzheimer’s disease” was first coined in 1906 when German psychiatrist Alois
Alzheimer described the case of a 51-year-old woman, Auguste D., who exhibited symptoms of
memory loss, confusion, and hallucinations. Upon her death, postmortem analysis revealed
abnormal protein deposits—now known as amyloid plaques and neurofibrillary tangles—in the
brain. These hallmark pathologies became the foundation for understanding
Alzheimer’[Link] the 20th century, diagnostic approaches relied primarily on clinical
evaluations—neuropsychological testing, cognitive scales like the Mini-Mental State Examination
(MMSE), and patient observation over time. However, such methods could only confirm

7
Alzheimer’s in its advanced stages when significant neuronal loss had already occurred. Medical
imaging advancements in the late 20th century—particularly Magnetic Resonance Imaging (MRI)
and Positron Emission Tomography (PET)—transformed neurological research by allowing
visualization of structural and functional brain changes. MRI scans, in particular, became
invaluable for identifying cortical atrophy, ventricular enlargement, and hippocampal shrinkage
associated with Alzheimer’s [Link] these technological advancements, manual
analysis of MRI scans remains subjective and time-intensive, often depending on expert
radiologists’ interpretation. Early or subtle structural changes may go undetected due to human
error or fatigue, emphasizing the need for computational approaches capable of consistent, large-
scale, and objective analysis.

MRI (Magnetic Resonance Imaging) provides high-resolution anatomical representations of the

brain, making it the preferred modality for Alzheimer’s analysis. The disease causes distinct
structural alterations, particularly in the hippocampus, amygdala, and cerebral cortex, leading to
measurable changes in brain volume and morphology. Early detection of these changes is vital
because pathological brain degeneration begins years before visible symptoms [Link]
Magnetic Resonance Imaging (MRI) has become one of the most essential non-invasive tools for
investigating neurodegeneration associated with Alzheimer’s disease. It allows detailed
visualization of brain anatomy, revealing tissue atrophy, cortical thinning, and hippocampal
shrinkage — biomarkers that correlate strongly with disease progression. Among these, medial
temporal lobe and hippocampal atrophy are widely accepted as key imaging indicators of early
Alzheimer’s. However, these changes are often subtle, spatially distributed, and highly variable
across individuals, making them difficult to detect through traditional image analysis or even fully
automated systems in heterogeneous, multi-site environments.

The complexity arises because MRI scans differ due to scanner types, imaging protocols, and
subject-specific factors. This heterogeneity introduces domain shifts, which can significantly
affect the performance and generalizability of deep learning models. Early studies applying
Convolutional Neural Networks (CNNs) to MRI-based Alzheimer classification achieved high
reported accuracies, but many relied on slice-level cross-validation or aggregate scan-level splits.
These evaluation strategies, although computationally convenient, can result in identity leakage,
where data from the same subject appears in both training and testing sets. This artificially inflates

8
model accuracy and undermines clinical validity, making the results unreliable in real-world
[Link] overcome these methodological flaws, recent studies emphasize subject-level
pipelines, where MRI data from each subject is carefully separated to prevent leakage and ensure
realistic generalization. These pipelines often involve aggregating information across multiple
axial slices rather than relying on single-slice predictions. The selected slices are standardized
through a series of preprocessing operations — including bias field correction, skull stripping,
spatial normalization, and intensity scaling — to ensure that the data fed into neural networks is
clean, consistent, and comparable across samples.

Traditional image processing methods—such as thresholding, segmentation, and statistical

analysis—require handcrafted feature extraction, which can be inconsistent across patients. Deep
learning, on the other hand, allows automatic feature learning from raw image pixels, ensuring
more robust pattern recognition. CNNs are especially well-suited for this task as they efficiently
capture local and spatial hierarchies of image features such as texture, edges, and intensity
[Link] CNNs excel in feature extraction, they often struggle with generalization and
overfitting when datasets are limited. Fully Connected Networks (DNNs), on the other hand,
perform better at classification when provided with high-quality feature representations. The 2D
Hybrid Neural Network proposed in this study combines these two architectures—leveraging
CNN layers for spatial feature extraction and dense layers for high-level decision-making.

This hybridization enables the model to:

 Capture detailed structural patterns from 2D MRI images

 Reduce information loss through optimized pooling and activation functions
 Improve classification accuracy through integrated dense networks
 Maintain computational efficiency suitable for real-world deployment

Such architectures balance deep learning performance with clinical practicality, offering scalable
solutions even in resource-constrained healthcare [Link] motivation behind this
research lies in addressing the growing need for automated, early, and accurate Alzheimer
detection systems. Early intervention significantly delays the onset of severe symptoms, providing
patients and caregivers with critical time to plan treatment and manage lifestyle adaptations.

9
However, traditional diagnostic methods—cognitive testing, behavioral observation, and manual
image review—remain inadequate for population-level [Link] developing countries like
India, where the ratio of neurologists to patients is alarmingly low and diagnostic facilities are
limited, AI-driven tools can democratize healthcare access. By using MRI-based automated
classification, even local healthcare centers could perform preliminary Alzheimer assessments,
referring only complex cases for specialized [Link], Alzheimer’s research
contributes to broader medical AI development, advancing image classification, pattern
recognition, and predictive analytics that can be adapted for other brain disorders such as
Parkinson’s disease, epilepsy, and brain tumors.

1.2 Identification of Clients and Needs/Tools

Artificial intelligence in healthcare is not developed in isolation—it exists to serve a specific group
of users, referred to as the “client,” who face challenges that technology can address. In the context
of Alzheimer’s disease detection, these clients encompass a wide range of stakeholders including
medical practitioners, diagnostic centers, healthcare institutions, caregivers, and research
organizations. Understanding their needs is essential to ensure that the proposed 2D Hybrid Neural
Network-based detection system is not only technically robust but also practically relevant and
ethically [Link] section outlines the identification of potential clients, their operational
context, the key problems they encounter in Alzheimer’s diagnosis, and the manner in which the
developed system addresses these needs [Link] and radiologists play a central
role in diagnosing neurodegenerative disorders. However, the increasing number of elderly
patients, limited availability of specialists, and the time-intensive process of manually analyzing
MRI scans create serious workflow challenges. In many hospitals, one radiologist may review
hundreds of MRI images daily, leading to fatigue, delayed reporting, and potential diagnostic
errors.

For clinicians, the proposed system serves as a computer-aided diagnostic (CAD) tool capable of:

 Automatically analyzing brain MRI scans to identify early-stage Alzheimer’s markers.

 Providing quantitative assessments such as probability scores for each classification (non-
demented, mildly demented, etc.).

10
  Offering visual heatmaps highlighting affected brain regions for easier interpretability.

This integration enhances clinical efficiency, reduces workload, and supports faster decision-
making. Most importantly, it acts as a second opinion system, assisting rather than replacing the
expertise of neurologists.

1.2.1 Problem Identification

Identifying client needs begins with understanding the core challenges in the current Alzheimer’s
diagnostic landscape.

1. Late Diagnosis: Most Alzheimer’s cases are diagnosed after significant brain
deterioration, reducing treatment effectiveness.
2. Human Dependency: Diagnosis heavily depends on expert interpretation, which is
subjective and time-intensive.
3. Data Overload: Hospitals generate thousands of MRI scans daily—manual analysis
of each image is infeasible.
4. Limited Resources: Rural and semi-urban healthcare centers often lack neurologists
or MRI specialists.
5. Cost and Accessibility: Advanced neuroimaging and specialist consultations
remain unaffordable for large segments of the population.
6. Technical Complexity: Existing 3D deep learning models require powerful GPUs
and expensive computational infrastructure, making them unsuitable for real-world
clinical settings.
7. Ethical and Interpretability Concerns: Clinicians are hesitant to trust black-box AI
systems without clear explanations for predictions.

These challenges collectively form the problem statement the proposed system seeks to
address:

To design a computationally efficient, interpretable, and accurate AI model capable of

detecting Alzheimer’s disease from MRI scans with minimal hardware requirements and
high reliability, suitable for real-world healthcare deployment.

11
Diagnostic centers and research institutions increasingly rely on AI-driven systems to automate
medical imaging tasks. These organizations require scalable solutions that can handle large
datasets, maintain consistency across multiple sites, and adapt to different imaging [Link]
research institutions such as Alzheimer’s Disease Neuroimaging Initiative (ADNI), National
Institute of Mental Health and Neurosciences (NIMHANS), or private AI-healthcare startups, the
proposed system addresses critical needs such as:

 Standardization: Ensuring consistent analysis across multi-site MRI data.

 Automation: Reducing manual annotation and feature extraction time.
 Reproducibility: Producing stable, repeatable diagnostic outcomes under controlled
parameters.
 Data Efficiency: Providing high accuracy even on limited training samples through hybrid
architecture and augmentation.

These features make the system a valuable research and diagnostic support tool for large-scale
Alzheimer’s studies and AI-powered health innovation [Link] third and most important
client group comprises patients and their caregivers. Alzheimer’s disease affects not only the
individual but the entire family ecosystem. Early detection enables timely treatment planning,
medication management, and emotional preparation for the progressive nature of the illness.

From the perspective of patients and families, the system fulfills the following needs:

 Early Intervention: Detecting Alzheimer’s before severe cognitive decline sets in.
 Accessibility: Providing screening capabilities in remote or under-resourced regions.
 Transparency: Offering explainable and visually interpretable reports.
 Affordability: Reducing dependency on costly high-end diagnostic procedures through
automation.

For communities with limited access to neurologists, such as rural India, an AI-based tool
integrated into local diagnostic centers can revolutionize early-stage screening and healthcare
accessibility.

12
The identification of clients and their needs forms the foundation of the Alzheimer detection
system’s design philosophy. Neurologists require intelligent assistance; researchers need
standardized, reproducible pipelines; patients and caregivers seek timely diagnosis and affordable
[Link] proposed 2D Hybrid Neural Network model effectively bridges these requirements,
combining computational innovation with clinical practicality. It recognizes that technology in
healthcare is not merely about automation—it is about augmentation: empowering medical
professionals, improving patient outcomes, and fostering a more inclusive and intelligent
healthcare [Link] meeting these multifaceted needs, this project not only addresses a
technological challenge but also fulfills a humanitarian mission—transforming how society
understands, detects, and manages Alzheimer’s disease in the age of artificial intelligence.

1.3 THE TECHNOLOGICAL REVOLUTION - From Traditional Diagnosis to

AI-driven Analysis

The 21st century has witnessed an extraordinary technological revolution that is transforming
every sphere of human life—from communication and industry to healthcare and scientific
research. Among these changes, one of the most impactful has been the integration of artificial
intelligence (AI) and machine learning (ML) into complex problem-solving domains once
considered the exclusive domain of human expertise. In the field of medicine, this revolution has
redefined diagnostic accuracy, personalized treatment, and predictive analysis. Within this
landscape, medical imaging—especially for neurodegenerative diseases such as Alzheimer’s—has
undergone a fundamental transformation driven by advances in computing power, algorithmic
design, and data [Link] journey from traditional medical diagnostics to AI-assisted
precision medicine has not been sudden but rather a gradual convergence of multiple technological
frontiers: digital imaging, computational modeling, cloud storage, and neural network-based
artificial intelligence. Together, these innovations have opened new possibilities for
understanding, detecting, and managing diseases that were previously beyond the scope of early
intervention.

Historically, medicine has been shaped by the tools available to visualize and interpret the human
body. From the invention of the stethoscope in 1816 to X-ray imaging in 1895 and MRI scanning
in the 1970s, each technological leap expanded the boundaries of medical understanding.

13
However, while these imaging technologies provided remarkable insights, their interpretation
relied heavily on human expertise. Radiologists and neurologists analyzed thousands of images
manually—an approach that was not only time-intensive but also prone to human error and
[Link] late 20th and early 21st centuries saw the emergence of digital imaging and
computer-aided diagnosis (CAD) systems, which marked the first attempts to computationally
assist medical professionals. These early systems used predefined algorithms and handcrafted
features to identify patterns, such as tumor edges or brain abnormalities. However, they lacked
adaptability and could not generalize well across diverse patient populations. The need for more
flexible, learning-based systems led to the rise of machine learning, where computers could learn
from data rather than follow fixed [Link] last decade has been revolutionary for AI applications
in healthcare. With exponential growth in computing capabilities, the availability of large medical
datasets, and the development of open-source AI frameworks, deep learning has become the
dominant paradigm in medical image analysis. Unlike traditional rule-based methods, AI systems
can autonomously learn from examples, identify subtle and nonlinear relationships, and
continuously improve through feedback loops.

In healthcare, AI’s contributions span across multiple domains:

 Diagnostics: Automated disease detection through image recognition (MRI, CT, X-rays).
 Predictive Analysis: Risk prediction models for cardiovascular, neurological, and
metabolic diseases.
 Drug Discovery: Accelerated molecular simulations and target identification.
 Personalized Treatment: Adaptive therapy based on genetic and behavioral data.
 Telemedicine and IoT: Real-time patient monitoring through wearable sensors integrated
with AI systems.

In the specific context of neurology, AI’s ability to analyze brain imaging data has become
particularly transformative. Alzheimer’s disease, with its slow progression and subtle biomarkers,
presents a unique challenge—requiring advanced computational models capable of detecting
patterns invisible to traditional image analysis methods. This challenge has catalyzed the
integration of AI-driven neuroimaging pipelines into both research and clinical workflows.

14
Deep learning represents a major shift from traditional AI paradigms. Instead of requiring manual
feature extraction, deep networks automatically learn hierarchical representations directly from
raw data. In image analysis, CNNs became the gold standard for their ability to capture spatial
hierarchies of features—from simple edges in early layers to complex anatomical structures in
deeper [Link] Alzheimer’s detection, CNNs analyze MRI scans to differentiate between normal
aging and pathological degeneration. Early studies demonstrated CNNs could achieve accuracies
exceeding 90% in classifying Alzheimer’s stages. However, while 3D CNNs offered a way to
process volumetric brain data holistically, they required immense computational resources,
making them impractical for large-scale or clinical settings. Conversely, 2D CNNs were
lightweight but often failed to capture inter-slice spatial [Link] led to a technological
evolution: the development of hybrid architectures—combining the efficiency of 2D CNNs with
additional layers or attention mechanisms to retain the contextual awareness of 3D models. These
hybrid networks mark the current frontier of medical AI, providing both computational efficiency
and diagnostic precision.

1.3.1 Hybrid Neural Networks: The Next Step in the Revolution

Hybrid neural networks integrate multiple deep learning paradigms to exploit their respective
advantages. In Alzheimer’s detection, a 2D Hybrid Neural Network typically combines:

 Convolutional Layers for spatial feature extraction from MRI slices.

 Dense or Fully Connected Layers for high-level classification.
 Dropout and Batch Normalization Layers for generalization and overfitting control.
 Attention Mechanisms to weight anatomically significant slices more heavily.

This design enables near-3D accuracy while maintaining low latency and memory efficiency—
making it ideal for real-world clinical deployment. Additionally, hybrid networks support subject-
level decision aggregation, where outputs from multiple 2D slices are combined to generate a
robust overall diagnosis.

The technological revolution here is twofold:

15
 1. Algorithmic Efficiency: Optimizing deep models to run effectively on standard hardware
without sacrificing performance.
2. Clinical Adaptability: Bridging the gap between research models and hospital-grade
diagnostic systems that can function in variable imaging conditions.

The technological revolution has ushered in a new era in medicine, where artificial intelligence
serves as a catalyst for progress in diagnosis, treatment, and patient care. In the context of
Alzheimer’s disease, this revolution has transformed neuroimaging from a manual interpretative
task into an intelligent, data-driven process capable of identifying microscopic brain changes years
before symptom onset.

The emergence of 2D Hybrid Neural Networks epitomizes this revolution’s essence—merging

efficiency with intelligence, clinical applicability with computational innovation. As this evolution
continues, the convergence of AI, cloud infrastructure, and ethical governance will not only
enhance diagnostic precision but also redefine the relationship between technology and human
health. The technological revolution in Alzheimer’s detection thus stands as both a scientific
breakthrough and a humanitarian milestone—bridging the gap between data and empathy, between
algorithms and healing.

1.4 Objectives and Scope of the Study

The global burden of Alzheimer’s disease is expanding at an alarming rate, creating an urgent
demand for intelligent diagnostic tools that can assist medical professionals in early detection and
intervention. The proposed project, “Alzheimer Detection via 2D Hybrid Neural Networks,” seeks
to harness the capabilities of artificial intelligence—particularly deep learning—to create a system
that can accurately and efficiently identify early signs of Alzheimer’s from brain MRI [Link]
chapter outlines the objectives, scope, and intended outcomes of the study, establishing a clear
roadmap for how the research contributes both scientifically and practically to the healthcare
[Link] objectives of this research define what the study is designed to accomplish. They
provide the guiding framework for methodology, experimentation, and evaluation. For a project
focused on Alzheimer’s detection using 2D hybrid neural networks, the objectives can be divided
into primary, secondary, and long-term goals.

16
1.4.1 Primary Objectives

The primary objectives represent the core aims—the fundamental problems this study seeks to
solve.

1. To design and implement a 2D Hybrid Neural Network model capable of accurately

classifying Alzheimer’s disease stages (Non-Demented, Very Mild Demented, Mild
Demented, and Moderate Demented) based on MRI brain images.
o The hybrid model will combine the feature extraction power of Convolutional
Neural Networks (CNNs) with the classification strength of Fully Connected
(Dense) layers, ensuring both accuracy and computational efficiency.
2. To develop an efficient, scalable, and explainable AI-based framework that can operate
effectively on standard hardware and provide real-time diagnostic support in clinical
environments.
o The system should maintain a balance between high accuracy and low latency,
making it suitable for hospital use and research deployment.
3. To integrate preprocessing and standardization techniques such as skull stripping, bias
correction, spatial normalization, and intensity scaling to ensure consistent, high-quality
input data and minimize domain variability across MRI datasets.
4. To evaluate the performance of the proposed model using rigorous statistical metrics,
including Accuracy, Precision, Recall, F1-score, and ROC-AUC, and to compare its
performance against existing deep learning models (such as standalone CNNs, 3D CNNs,
and transfer learning architectures).
5. To generate interpretable results through visualization tools, including heatmaps and
attention maps, allowing medical professionals to understand the model’s predictions and
gain trust in its diagnostic suggestions.

1.4.2 Secondary Objectives

The secondary objectives extend beyond the technical design and explore the broader implications
and optimization of the study.

17
 1. To enhance the reliability and generalization of the model by applying data augmentation,
cross-validation, and subject-level split strategies to prevent overfitting and data leakage.
2. To develop a lightweight deployment-ready version of the model that can be integrated
into desktop or web-based interfaces, enabling doctors, radiologists, or research assistants
to use it conveniently in their diagnostic workflows.
3. To analyze the comparative performance of 2D and 3D CNN approaches, highlighting the
trade-offs between computational complexity, accuracy, and real-world deployability.
4. To establish a foundation for multimodal disease detection systems, where the same
architecture could be adapted for other neurodegenerative disorders such as Parkinson’s
disease, dementia with Lewy bodies, or frontotemporal degeneration.
5. To explore the ethical and data governance considerations of using AI in medical
diagnostics, particularly concerning privacy, explainability, and accountability.

1.4.3 Long-Term Objectives

While this project primarily focuses on Alzheimer’s detection, the long-term vision is to contribute
to the ongoing evolution of AI-assisted healthcare ecosystems. The study aims:

1. To demonstrate that deep learning, when combined with proper preprocessing and
hybridization, can bridge the gap between computational research and clinical
implementation.
2. To contribute an open, reproducible model pipeline that can serve as a benchmark for future
Alzheimer’s detection research.
3. To inspire the integration of AI-based tools into community healthcare, where early
screening could be conducted even in resource-limited environments through cloud-based
or IoT-connected systems.
4. To advance the global effort toward precision medicine, where individualized, data-driven
analysis aids early detection, prediction, and prevention of neurodegenerative diseases.

The scope of the study defines the boundaries within which the research operates—what it
includes, what it excludes, and the extent to which its outcomes can be generalized. For “Alzheimer

18
Detection via 2D Hybrid Neural Networks,” the scope encompasses the technical, medical, and
practical dimensions of applying deep learning to neuroimaging.

From a medical and diagnostic perspective, the study focuses on developing an AI-based decision
support tool rather than a replacement for clinical judgment. Target Condition Alzheimer’s disease
and its early-stage detection through neuroimaging. Detection Focus: Identification of brain
atrophy patterns in regions such as the hippocampus, medial temporal lobe, and cerebral cortex.
Clinical Role: Providing radiologists and neurologists with a supplementary diagnostic tool that
enhances their accuracy and efficiency. Intended Use: Clinical research, academic studies, and
diagnostic assistance in hospitals or specialized memory care units.

The research scope extends to exploring deep learning architectures and hybrid design principles.
It includes:

 Comparative study between standard CNNs, 3D CNNs, and hybrid 2D CNN-DNN models.
 Analysis of the effect of preprocessing standardization on model accuracy.
 Investigation of dataset balancing methods such as Synthetic Minority Over-sampling
Technique (SMOTE) for class imbalance.
 Evaluation of interpretability enhancements through attention mechanisms.

The objectives and scope of this study collectively define its academic and practical trajectory.
The primary objective—developing an efficient, accurate, and interpretable 2D Hybrid Neural
Network—is aligned with the larger goal of democratizing AI-driven healthcare. The project’s
scope, though bounded by computational and dataset constraints, ensures meaningful contributions
to the field of neuroimaging and clinical [Link] clearly identifying the model’s purpose,
operational limits, and deliverables, this chapter establishes the foundation for the methodological
and experimental design that follows. The study’s long-term vision extends beyond Alzheimer’s
detection—toward a future where deep learning systems become trusted companions in the global
fight against neurodegenerative diseases.

19
 CHAPTER 2.
LITERATURE REVIEW

2.1 Timeline of the Reported Problem

Understanding the evolution of Alzheimer’s disease research and the technological progression in
its detection is essential to contextualize the present study. The history of Alzheimer’s diagnosis
spans more than a century, encompassing discoveries in medicine, neurobiology, imaging, and,
more recently, artificial intelligence. This timeline illustrates how the problem of Alzheimer’s
detection evolved from manual observation to advanced hybrid neural network models that are
redefining diagnostic precision.

Early 1900s – The Discovery of Alzheimer’s Disease

The story of Alzheimer’s disease begins in 1906, when German psychiatrist and neuropathologist
Dr. Alois Alzheimer presented his findings on a patient named Auguste D., who exhibited
progressive memory loss, confusion, and behavioural disturbances. Postmortem examination of
her brain revealed amyloid plaques and neurofibrillary tangles, now recognized as the pathological
hallmark of the disease This discovery marked the birth of Alzheimer’s disease as a medical
condition, although it remained poorly understood for several decades. During this early period,
diagnosis was entirely clinical and observational, relying on behavioural and memory assessments
rather than imaging or biological markers.

1930s–1960s – The Neuropathological Phase

Throughout the mid-20th century, Alzheimer’s disease was often classified as a rare form of
“senile dementia.” Neuropathological studies during this period focused on brain tissue analysis,
primarily through autopsies. Researchers identified significant neuronal loss in the hippocampus
However, there was no method to detect these changes in living patients, making early diagnosis
impossible. The problem remained primarily postmortem identification, and effective treatment
options were non-existent.

20
1970s–1980s – The Neuroimaging Revolution

The 1970s brought the first major technological leap in Alzheimer’s research with the advent of
Computed Tomography (CT) and Magnetic Resonance Imaging (MRI). These imaging techniques
enabled the non-invasive visualization of brain structures, allowing clinicians to identify pattern
By the late 1980s, MRI had become the gold standard for neurological imaging, offering higher
resolution and better soft tissue contrast than CT scans. Researchers began using MRI to detect
hippocampal atrophy — one of the earliest visible indicators of Alzheimer’s [Link],
even with MRI, analysis was primarily manual, relying on the radiologist’s visual interpretation.
Differences between scanners, image quality, and observer bias made consistent diagnosis
difficult. The need for automated, quantitative, and objective methods became increasingly
apparent.

1990s – The Emergence of Computational Neuroimaging

The 1990s marked the beginning of computational neuroscience, where statistical and
mathematical tools were used to analyze brain images. Researchers started employing image
segmentation, volumetric analysis, and voxel-based morphometry to quantify brain structure
[Link] tools such as SPM (Statistical Parametric Mapping) and FSL (FMRIB Software
Library) were introduced, enabling standardized analysis [Link] these advancements,
early computational approaches depended heavily on handcrafted features and manual region-of-
interest (ROI) selection. The process required expert domain knowledge and suffered from
scalability issues. Automated methods capable of detecting complex, subtle patterns in MRI data
were still in their infancy.

2000s – The Rise of Machine Learning in Medicine

The early 2000s introduced machine learning (ML) as a new paradigm in medical image analysis.
Algorithms such as Support Vector Machines (SVM), Random Forests (RF), Decision Trees, and
k-Nearest Neighbors (kNN) were applied to Alzheimer’s detection tasks.
These models used statistical features derived from MRI scans—such as gray matter volume,
cortical thickness, and texture—to classify subjects as healthy or demented.

21
During this era, several landmark studies demonstrated the potential of ML-based Alzheimer’s
classification:

Fan et al. (2008) used SVMs to classify Alzheimer’s patients based on regional brain
volumes, achieving accuracies above 80%.Davatzikos et al. (2009) introduced pattern
classification methods for identifying early-stage Alzheimer’s from structural MRI data.

While these methods improved diagnostic consistency, they still relied on manual feature
extraction—a major limitation that prevented fully automated systems. The transition to deep
learning, where models learn features directly from raw data, would soon revolutionize the field.

2010–2020 – The Deep Learning Breakthrough

The decade of 2010 marked the true technological revolution in Alzheimer’s detection, driven by
the success of Deep Learning (DL) and Convolutional Neural Networks (CNNs).
The seminal success of AlexNet (2012) in the ImageNet competition demonstrated that CNNs
could automatically learn hierarchical features from raw images, outperforming all traditional ML
methods. This innovation soon extended into medical imaging.

Researchers began experimenting with CNNs for Alzheimer’s diagnosis using MRI and PET
scansCNNs eliminated the need for handcrafted features, enabling end-to-end learning directly
from imaging data. However, their performance was constrained by limited labeled datasets and
high computational demands, especially for 3D brain volumes.

2020–2022 – The Rise of Hybrid and Attention-Based Neural Networks

Recent years have seen a rapid shift toward hybrid neural architectures that combine the eff
The 2D Hybrid Neural Network approach uses 2D convolutional layers to process MRI slices
efficiently and aggregates slice-level predictions using dense layers or attention mechanisms to
produce subject-level classifications.

 Valliani and Soni (2020) introduced a hybrid CNN-LSTM model that captured both
spatial and temporal dependencies in MRI sequences.

22
  Lu et al. (2022) proposed a lightweight hybrid model that achieved comparable accuracy
to 3D CNNs while reducing computational cost by nearly 60%.

These developments mark a crucial turning point—making Alzheimer detection systems not only
accurate but also scalable and deployable in real clinical workflows.

2023–Present – Toward Explainable and Multimodal AI

The latest research trends emphasize explainability, fairness, and multimodal integration in AI-
Hybrid models are now being extended with Explainable AI (XAI) tools such as Grad-CAM and
Layer-wise Relevance Propagation (LRP) to visualize decision-making processes and build
clinical trust. Furthermore, researchers are combining MRI, PET, fMRI, and genetic data (e.g.,
APOE genotyping) to build holistic disease models that predict not only the presence but also the
progression of Alzheimer’s.

In addition, frameworks like federated learning and cloud-based AI systems are being developed
to address data privacy challenges and enable collaborative, large-scale model training across
hospitals without sharing sensitive patient data. The current frontier involves 2D and 3D hybrid
models with transformer-based architectures, capable of capturing both spatial and semantic
information, signifying the most advanced stage of Alzheimer’s detection research to date.

The evolution of Alzheimer’s detection has mirrored the broader technological transformation in
medicine. From early microscopic discoveries to advanced hybrid neural architectures, each stage
has brought greater precision, automation, and understanding. Despite over a century of progress,
the fundamental challenge remains—detecting Alzheimer’s at the earliest possible stage with
accuracy, efficiency, and interpretability. The emergence of 2D Hybrid Neural Networks
represents a pivotal step in this timeline: an approach that combines computational feasibility with
diagnostic depth, bridging the gap between academic research and real-world clinical application.

This historical context not only justifies the need for the present study but also highlights how
modern AI continues to push the boundaries of neuroscience, transforming Alzheimer detection
into a truly data-driven discipline.

23
2.2 Bibliometric Analysis

The rapid advancement of artificial intelligence (AI), particularly deep learning, has led to an
exponential increase in research publications focused on Alzheimer’s disease (AD) detection and
classification using neuroimaging data. Bibliometric analysis provides a structured approach to
examine this growing body of literature, identify key research trends, evaluate the most influential
contributions, and recognize the technological and methodological evolution of the [Link]
section presents a bibliometric analysis of scholarly publications related to Alzheimer’s disease
detection using machine learning and deep learning approaches. The analysis considers aspects
such as publication growth over time, most cited works, dominant research methodologies,
commonly used datasets, and emerging directions in the intersection of neuroscience and
computational intelligence.

Bibliometric analysis is a quantitative evaluation of research output within a specific domain. In

the context of Alzheimer’s disease detection, it serves to:

 Understand how research focus has evolved from traditional neuroimaging to AI-driven
diagnostic techniques.
 Identify the most commonly used datasets, frameworks, and models.
 Evaluate which research institutions and countries have made significant contributions.
 Detect emerging trends, research gaps, and underexplored areas that motivate current
studies—such as the need for efficient and explainable hybrid models.

This methodical evaluation thus helps contextualize the present project, “Alzheimer Detection
via 2D Hybrid Neural Networks,” within global research efforts.

The bibliometric data highlights several leading journals and conferences that consistently publish
influential Alzheimer detection research:

24
Over the last two decades, Alzheimer’s disease has remained one of the most widely studied
neurological disorders, not only for its medical relevance but also for the computational challenges
it poses. The disease’s complex manifestation in brain structures, combined with the variability of
MRI data, has made it a rich testing ground for developing and validating AI algorithms.
Consequently, the number of academic publications addressing Alzheimer’s detection through
computational methods has grown exponentially, particularly after 2015. This surge coincides with
the global adoption of deep learning frameworks such as TensorFlow and PyTorch, which made
high-performance neural networks accessible to researchers beyond the core fields of computer
science.

During the early 2000s, the research landscape was dominated by statistical and machine learning
approaches—Support Vector Machines, Random Forests, and Principal Component Analysis were
the preferred tools for identifying structural variations in the brain. These models relied heavily on
handcrafted features extracted from specific regions of interest, such as the hippocampus and
medial temporal lobes. Studies during this period were technically limited but conceptually
significant, as they laid the groundwork for automated analysis of neuroimaging data.

However, the bibliometric pattern shows a marked shift around 2012, when convolutional neural
networks (CNNs) began entering medical image analysis. Inspired by their unprecedented success
in object recognition and visual classification tasks, researchers adapted CNN architectures for 2D
25
and 3D MRI data. The impact of this transition is clearly visible in publication metrics: the number
of Alzheimer detection studies using CNNs nearly tripled between 2015 and 2019. Deep learning
had officially replaced traditional machine learning as the dominant paradigm.

A closer reading of the literature reveals that this shift was not only technological but
methodological. Earlier models required human-defined features—texture descriptors, edge
patterns, or shape measurements—while CNNs could automatically learn spatial hierarchies
directly from raw MRI data. This removed human bias from the feature engineering process and
enabled more robust, end-to-end learning frameworks. Bibliometric analysis during this phase
identifies a set of highly cited foundational papers, such as those by Suk et al. (2014) and Payan
and Montana (2015), which introduced multimodal and 3D deep learning models for Alzheimer’s
classification. These works collectively signified the beginning of deep neural network
applications in clinical imaging.

The next significant wave of research, between 2016 and 2020, focused on optimizing these
models for higher accuracy and broader generalization. Three-dimensional CNNs became
increasingly common, as they allowed the model to capture inter-slice relationships within
volumetric MRI data. However, bibliometric data also indicates a rising concern about
computational overhead and reproducibility during this period. Many 3D models, though accurate
in controlled experiments, required high-end GPUs and massive memory, making them
impractical for real-time or large-scale clinical deployment. As a result, research attention began
shifting toward hybrid architectures—those combining the efficiency of 2D networks with the
contextual strength of 3D representations.

From 2020 onward, a new generation of studies emerged emphasizing lightweight, interpretable,
and clinically viable AI systems. These models integrated attention mechanisms, residual learning,
and explainable AI techniques like Grad-CAM to make neural decisions more transparent to
medical practitioners. A bibliometric review of publications between 2020 and 2024 shows
frequent recurrence of keywords such as “hybrid CNN,” “attention model,” “transfer learning,”
and “explainable AI.” This demonstrates a maturing of the field, as researchers began to move
beyond raw performance metrics toward trust, usability, and ethical compliance.

26
In this maturing stage, the 2D Hybrid Neural Network approach—like the one proposed in the
present study—emerged as a crucial research direction. Bibliometric patterns confirm that 2D
models have gained renewed attention due to their lower hardware requirements and adaptability
for resource-constrained healthcare environments. By aggregating information across multiple
MRI slices and integrating attention mechanisms, hybrid 2D architectures achieve accuracies
comparable to 3D CNNs while consuming a fraction of the computational resources. This
convergence of efficiency and interpretability marks the current frontier of Alzheimer detection
[Link] important insight from bibliometric evaluation concerns dataset utilization.
Nearly all studies in the deep learning era depend heavily on standardized open-access datasets
such as the Alzheimer’s Disease Neuroimaging Initiative (ADNI), OASIS, and AIBL. These
datasets have become the de facto benchmarks for training and testing deep learning models. The
ADNI dataset, in particular, appears most frequently across publications, with over half of all deep
learning-based Alzheimer studies citing it as their primary data source. This dependency ensures
comparability of results across research groups but also exposes the limitation of model
generalization. Bibliometric reviews highlight that very few studies have validated their models
on diverse, real-world clinical datasets. This gap points to the urgent need for adaptive,
generalizable architectures—further justifying the rationale for lightweight hybrid networks that
can be retrained on local data.

Despite its enormous progress, the bibliometric synthesis uncovers several persisting weaknesses
within the body of literature. Many published studies continue to use limited datasets or lack
external validation. Cross-validation methods are sometimes flawed, leading to potential data
leakage and overestimated performance. Moreover, although explainable AI is an active area of
discussion, only a small fraction of papers provide truly interpretable visualizations of neural
network decisions. This lack of transparency reduces the clinical reliability of otherwise powerful
models. Another identified gap concerns real-world integration; few studies move beyond
laboratory evaluation to deployable systems within hospitals or diagnostic [Link], the
bibliometric review serves not merely as a record of past achievements but as a diagnostic map of
the field’s shortcomings and opportunities. It emphasizes that the next phase of Alzheimer
detection research should focus on creating computationally efficient, transparent, and clinically
adaptable models—an exact reflection of the objectives pursued in the present study. The 2D

27
Hybrid Neural Network proposed in this project directly responds to these unmet needs, offering
a system that is lightweight enough for real-world use yet powerful enough to maintain diagnostic
accuracy comparable to more complex architectures.

In summary, the bibliometric analysis demonstrates that Alzheimer’s detection research has
progressed through distinct yet overlapping phases—from clinical observation and feature-based
modeling to deep learning and, finally, to explainable hybrid architectures. Each stage addressed
the limitations of its predecessor but introduced new challenges in complexity, interpretability, or
accessibility. The convergence of these insights underlines the need for balanced models that are
accurate, efficient, and interpretable. Within this continuum, the proposed hybrid approach aligns
precisely with the direction modern research is taking—representing both an evolution of deep
learning and a step toward the practical realization of AI in healthcare.

2.3 Problem Definition

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder that leads to irreversible

memory loss, cognitive decline, and behavioral impairment. The global burden of Alzheimer’s is
immense, affecting over 55 million people worldwide, with numbers projected to nearly double
by 2050. Despite decades of research, early and reliable diagnosis remains one of the most pressing
unsolved challenges in neuroscience. Conventional diagnostic practices—such as cognitive
assessments, neurological evaluations, and manual MRI interpretation—are subjective, time-
consuming, and prone to human [Link] Magnetic Resonance Imaging (MRI) provides
detailed structural information about the brain, its interpretation demands specialized expertise.
Radiologists must manually assess subtle atrophic patterns in regions such as the hippocampus and
medial temporal lobes, which are difficult to quantify, especially in early disease stages. Moreover,
inter-observer variability and scanner-dependent differences make manual evaluation inconsistent
and unreliable for population-scale screening.

The integration of machine learning (ML) and deep learning (DL) into Alzheimer’s research
promised a paradigm shift toward objective, data-driven diagnostics. Early ML approaches, based
on statistical classifiers such as Support Vector Machines (SVM) and Random Forests, achieved
moderate success but required extensive handcrafted feature extraction. This dependency on

28
domain expertise limited their generalization ability and prevented large-scale clinical
[Link] the rise of deep learning—especially Convolutional Neural Networks (CNNs)—
models began automatically learning complex imaging features directly from MRI data. Studies
from 2015 onward demonstrated significant performance improvements using CNNs and 3D
neural architectures. However, as highlighted in the bibliometric analysis, these advancements
brought new challenges:

 Computational inefficiency: 3D CNN models demand large GPU memory and high
processing power, making them unsuitable for smaller healthcare centers.
 Overfitting and poor generalization: Many models perform well on benchmark datasets
but fail on new data due to limited diversity and data leakage issues.
 Lack of explainability: Deep learning predictions often operate as “black boxes,” offering
little insight into the biological reasoning behind classification outcomes.
 Data imbalance and limited availability: MRI datasets for Alzheimer’s are relatively
small and geographically biased, further hindering model robustness.
 Clinical inadaptability: Many architectures are too complex or resource-heavy to
integrate into existing hospital systems or local diagnostic workflows.

These challenges define the core research problem: while existing deep learning models have
achieved high theoretical accuracy, they remain computationally heavy, difficult to interpret, and
insufficiently validated for real-world diagnostic use.

2.4 Goals & Objectives

In response to the defined problem, the overarching goal of this project is to design and implement
an AI-driven diagnostic framework that leverages 2D Hybrid Neural Networks to detect
Alzheimer’s disease efficiently and accurately from MRI brain images. The model aims to blend
the computational simplicity of 2D CNNs with the context-awareness of hybrid learning
structures, producing a system that is fast, interpretable, and adaptable for real-time applications.

At its core, the project seeks to contribute to the ongoing transformation of medical imaging from
a manually interpretive field to a data-driven, automated discipline — one where artificial
intelligence serves as a reliable partner to clinicians rather than a replacement.

29
The specific goals of this study can be conceptually summarized as follows:

1. To design a 2D hybrid neural architecture that extracts meaningful spatial and structural
features from MRI slices and aggregates them to form a robust subject-level diagnosis.
2. To improve computational efficiency and enable deployment on standard computing
systems without reliance on expensive GPU resources.
3. To implement an attention-based or explainable module that highlights key anatomical
regions influencing the model’s predictions, thereby increasing transparency and trust
among clinicians.
4. To establish a standardized preprocessing pipeline (including bias correction, skull
stripping, and normalization) that ensures consistent data quality and fair evaluation across
heterogeneous datasets.
5. To validate the proposed model rigorously using widely recognized metrics — accuracy,
precision, recall, F1-score, ROC-AUC — ensuring statistical reliability and
reproducibility.
6. To align the system’s design philosophy with ethical AI principles — fairness,
interpretability, privacy, and accountability — which are essential for medical applications.

Ultimately, the goal is to demonstrate that lightweight and interpretable deep learning architectures
can achieve near state-of-the-art performance while maintaining the practicality required for
clinical adoption.

The literature and bibliometric analyses reveal specific deficiencies in current approaches, which
translate into actionable objectives for this study. These objectives ensure that the research moves
beyond theoretical contribution toward tangible, applied [Link] first key objective is to bridge
the computational gap between 2D and 3D convolutional models. While 3D CNNs capture
volumetric information effectively, they require extensive processing power and training time. The
proposed hybrid model aims to replicate the contextual depth of 3D models by combining
information across multiple 2D slices, thus achieving a balance between performance and
efficiency.

30
The second objective is to enhance interpretability through attention and visualization
mechanisms. As seen in the reviewed studies, the lack of explainability remains a barrier to AI
acceptance in medicine. By incorporating Grad-CAM-based visualization, the system can
highlight the regions contributing most to classification decisions, helping doctors correlate
algorithmic findings with clinical observations. The third objective focuses on standardization and
reproducibility. Many existing models use inconsistent preprocessing or biased data partitions,
leading to inflated results. This study emphasizes rigorous, leakage-proof subject-level data
splitting and a reproducible preprocessing pipeline to ensure that performance metrics genuinely
reflect the model’s real-world potential.

Another important objective is accessibility. The project’s design is intentionally lightweight,

enabling deployment in small hospitals, diagnostic laboratories, or even cloud-based telemedicine
systems. This goal aligns with the global movement toward democratizing AI — ensuring that
advanced diagnostic technology is not confined to elite research centers but accessible to
communities [Link], the study aims to set the foundation for extensibility. While the
current model focuses on Alzheimer’s detection, the underlying architecture can be adapted to
other neurodegenerative diseases, such as Parkinson’s or multiple sclerosis, or extended to
multimodal imaging involving PET or fMRI [Link], these objectives translate the
theoretical understanding of existing research into a practical roadmap for innovation, ensuring
that the study contributes both academically and societally. The defined goals and objectives
therefore revolve around building a system that is accurate, lightweight, interpretable, and
reproducible — in other words, a system that bridges the gap between research prototypes and
real-world medical solutions.

This project not only seeks to advance the scientific field of computational neuroscience but also
aspires to make AI a trustworthy and accessible diagnostic companion in the fight against
Alzheimer’s disease. The realization of these goals will mark a step forward in the ongoing
transformation of medicine — from reactive care to predictive, data-driven intelligence — where
every scan contributes to saving both time and memory.

31
 CHAPTER 3.
DESIGN FLOW/PROCESS

3.1 Data Collection and Preprocessing

The development of an effective Alzheimer’s detection system depends critically on the careful
evaluation and selection of both technical specifications and data features. Every design choice —
from dataset type to feature extraction strategy and model configuration — directly influences the
accuracy, efficiency, and interpretability of the final neural network. This section provides an in-
depth discussion of the evaluation process used to identify the most suitable features, parameters,
and specifications for building a 2D Hybrid Neural Network optimized for Alzheimer’s disease
classification using MRI images.

The goal of this stage is twofold:

1. To ensure that the model inputs (data features) capture the most diagnostically relevant
aspects of Alzheimer’s pathology.
2. To configure the system architecture and computational environment in a way that ensures
reliability, scalability, and real-world deploy ability.

The following analysis draws upon insights from prior literature, experimental validation, and
medical domain considerations. In deep learning, model performance is heavily dependent on the
quality and representation of the input data. For neuroimaging applications like Alzheimer’s
detection, the features of interest are subtle and spatially distributed — small variations in gray
matter density, cortical thickness, or hippocampal volume can significantly influence the
diagnosis. Consequently, the evaluation and selection of features must be driven by both biological
relevance and computational [Link] a medical perspective, Alzheimer’s disease is
characterized by progressive neurodegeneration in the hippocampus, temporal lobes, and cerebral
cortex, leading to structural brain atrophy visible in MRI scans. These physical changes form the
foundation for selecting features and slices most informative for model training. From a
computational standpoint, the challenge lies in representing these 3D brain changes effectively
using 2D convolutional operations, without losing contextual depth. The hybrid model proposed

32
in this study addresses this challenge through intelligent feature extraction and slice aggregation
strategies.

Before feature extraction, the first specification requiring evaluation is the dataset. The model’s
success depends on the diversity, size, and representativeness of the training data. After reviewing
various public datasets, including ADNI, OASIS, and Kaggle Alzheimer MRI datasets, the Kaggle
Alzheimer MRI Dataset was chosen as the primary training source due to its accessibility,
standardized 2D format, and predefined labeling categories: Non-Demented, Very Mild Demented,
Mild Demented, and Moderate Demented.

The selection criteria were based on the following aspects:

1. Data Diversity: The dataset contains a balanced representation of all disease stages,
minimizing the class imbalance problem common in clinical data.
2. Quality and Preprocessing Level: MRI images are already skull-stripped and normalized,
ensuring high visual quality and consistent intensity distribution.
3. Compatibility: As the study focuses on 2D CNN-based architectures, this dataset
eliminates the need for 3D volumetric reconstruction, reducing computational complexity.

However, to improve generalization and model robustness, preprocessing and data augmentation
techniques are further applied, ensuring the features extracted from these images remain medically
meaningful and computationally stable. Feature extraction in deep learning is governed by the
architecture specifications of the neural network. The hybrid 2D CNN model designed for this
study was carefully evaluated to ensure optimal depth, kernel size, activation functions, and
regularization methods.

The evaluation process involved analysing how different configurations affect feature quality and
model performance:

 Convolutional Layers: Three to four convolutional blocks were found optimal, balancing
expressiveness with computational efficiency. Each layer captures progressively abstract
features — from edges and textures to high-level structural differences. A kernel size of
3×3 ensures fine-grained feature extraction suitable for MRI patterns.

33
  Pooling Layers: Max-pooling layers with 2×2 kernels reduce spatial dimensions while
retaining dominant features. This not only controls model size but also introduces
translational invariance — an important property when MRI slices vary slightly in
orientation.
 Activation Functions: The Rectified Linear Unit (ReLU) activation function was selected
for its ability to introduce non-linearity while maintaining efficient gradient propagation,
enabling deeper models to converge faster.
 Dropout and Batch Normalization: Regularization is essential to avoid overfitting on
limited medical datasets. Dropout layers randomly deactivate neurons during training,
while batch normalization stabilizes the learning process and accelerates convergence.

The performance and feasibility of any deep learning system are heavily dependent on the
computational resources available for its development, training, and deployment. In the case of
Alzheimer’s detection through MRI analysis, the system must be capable of handling large
volumes of high-resolution medical images while maintaining reasonable training times and
consistent performance. The proposed 2D Hybrid Neural Network model is specifically designed
to operate efficiently within standard computational environments, emphasizing accessibility and
resource optimization rather than dependence on high-end specialized [Link] of the core
objectives of this research is computational efficiency — ensuring that the system performs
reliably on common setups available in academic laboratories or medium-scale diagnostic centers.
Many existing studies employ resource-intensive 3D Convolutional Neural Networks (CNNs) that
require extensive GPU memory and processing power, often making them impractical for real-
world healthcare environments. In contrast, the hybrid 2D architecture developed in this project
focuses on achieving high diagnostic accuracy using manageable computational specifications.

The experiments and model training were conducted on a workstation configured with an Intel i7
or Ryzen 7 processor, paired with 16–32 GB of RAM. For hardware acceleration during training,
an NVIDIA GPU such as the GTX 1660 Ti or RTX 3060 was employed. These GPUs provide a
balanced trade-off between computational capability and affordability, supporting large batch
processing and fast gradient updates without the excessive energy and cost overheads associated
with high-end enterprise GPUs like the Tesla or A100 series.

34
From a software perspective, the entire model pipeline was built within the Python 3.8+
environment due to its robustness, extensive libraries, and strong community support in the field
of machine learning and data science. The deep learning framework TensorFlow 2.x was used
alongside Keras, providing a modular and user-friendly interface for model construction and
optimization. Data preprocessing, augmentation, and manipulation were performed using NumPy
and Pandas, while Matplotlib and Seaborn were used for data visualization and performance
analysis. This selection of frameworks was made deliberately to ensure compatibility, scalability,
and reproducibility across multiple systems and operating environments.

Evaluation of Performance Metrics and Feature Validation

While computational efficiency defines the practical feasibility of a system, performance

evaluation metrics determine its scientific and diagnostic credibility. In the context of Alzheimer’s
detection, where medical decisions can directly affect patient outcomes, evaluation cannot rely
solely on overall accuracy. Instead, it must encompass a multidimensional assessment that reflects
how well the model identifies true cases, avoids false diagnoses, and maintains reliability across
different classes of disease [Link] process of performance evaluation serves two essential
purposes. First, it validates the effectiveness of the selected features and architectural
configurations — confirming that the CNN’s learned representations genuinely capture
meaningful patterns related to Alzheimer’s pathology. Second, it ensures that the system meets
clinical reliability standards, providing diagnostic outputs that physicians can trust in practice.

Given the critical nature of Alzheimer’s detection, this study employs a comprehensive set of
evaluation metrics, including Accuracy, Precision, Recall (Sensitivity), F1-Score, and ROC-AUC.
Each of these metrics measures a distinct dimension of performance, collectively providing a
nuanced picture of the model’s diagnostic [Link] is the most straightforward and
commonly reported measure. It quantifies the proportion of correctly classified MRI scans among
all predictions. While high accuracy indicates good overall performance, it can be misleading when
class distributions are imbalanced — a common occurrence in medical datasets where healthy
subjects often outnumber patients. Therefore, accuracy is interpreted here as a general indicator

35
rather than a definitive measure of clinical [Link] complement this, Precision evaluates the
reliability of positive predictions. In this study, it measures the proportion of scans identified as
Alzheimer’s that are truly diseased. High precision indicates that the model minimizes false
positives — crucial for preventing unnecessary anxiety and interventions in healthy individuals.

3.2 Data Collection and Preprocessing

The success of any deep learning-based medical diagnosis system is fundamentally determined by
the quality, diversity, and representativeness of the data on which it is trained. In the context of
Alzheimer’s disease detection, the use of neuroimaging data such as Magnetic Resonance Imaging
(MRI) provides a rich and non-invasive source of structural information about the brain. However,
the raw MRI data acquired from various sources is often inconsistent, noisy, and high-dimensional,
necessitating systematic preprocessing before it can be used effectively in neural network
[Link] section presents a comprehensive theoretical overview of the processes involved
in data collection and preprocessing for the proposed 2D Hybrid Neural Network model. It covers
the rationale for dataset selection, data characteristics, preprocessing techniques applied to ensure
data consistency, and the transformation of raw medical images into clean, structured, and
standardized input suitable for efficient model training and reliable classification.

The data used in this study was collected from publicly available and medically validated sources
specializing in Alzheimer’s disease neuroimaging research. Among various open-access
repositories such as the Alzheimer’s Disease Neuroimaging Initiative (ADNI), OASIS (Open
Access Series of Imaging Studies), and AIBL (Australian Imaging, Biomarkers & Lifestyle Study),
the Kaggle Alzheimer MRI Dataset was chosen as the primary dataset for the present research due
to its structured labeling, uniform imaging resolution, and ease of integration with 2D
convolutional architectures.

The selected dataset contains MRI slices of the human brain categorized into four diagnostic
stages:

1. Non-Demented (healthy individuals)

36
 2. Very Mild Demented
3. Mild Demented
4. Moderate Demented

This categorization facilitates multi-class classification, allowing the model not only to distinguish
between healthy and diseased states but also to identify the progressive stages of cognitive decline.
The dataset comprises thousands of grayscale MRI images derived from different patients, each
slice representing a cross-sectional view of the brain. The use of two-dimensional slices, as
opposed to full 3D volumetric scans, is intentional — it aligns with the goal of creating a
computationally efficient and resource-friendly model suitable for real-world deployment.

The images in this dataset are pre-labelled and quality checked by domain experts. This ensures
that the ground truth used during training is reliable, which is critical for minimizing biases in
model learning. Nevertheless, as with most real-world medical datasets, even high-quality MRI
images require substantial preprocessing to standardize intensity values, remove non-brain tissues,
and mitigate scanner-related variations that could otherwise distort the learning process. RI data
differs significantly from conventional image data used in computer vision tasks. Unlike natural
images, MRI scans represent volumetric anatomical structures captured using magnetic resonance
signals rather than optical reflections. Each pixel intensity corresponds to a tissue density or signal
intensity rather than a colour gradient. Consequently, these images lack uniform contrast and may
vary depending on scanning parameters, Therefore, even minor variations in brightness, noise, or
alignment across MRI slices can obscure diagnostically significant patterns. This makes
preprocessing a vital stage to ensure that the model focuses on medically relevant information
rather than irrelevant scanner artifacts or random noise.

Spatial Alignment and Orientation Normalization

Even after resizing and cropping, MRI images may vary in orientation due to differences in head
positioning during scanning. Misalignment can significantly affect the performance of
convolutional networks, which rely on spatial consistency across training samples. Therefore, an
alignment procedure known as spatial normalization is [Link] this step, MRI slices are aligned
to a standardized coordinate system — typically the Montreal Neurological Institute (MNI) space

37
— ensuring that all brain regions correspond spatially across subjects. This alignment enables the
network to learn spatially consistent features from homologous brain regions, improving
classification robustness and interpretability.

Data Augmentation and Synthetic Diversity

Deep learning models thrive on large and diverse datasets. However, collecting and labelling
extensive medical data is time-consuming and ethically constrained. To compensate for limited
data availability, this project employs data augmentation techniques to synthetically increase
dataset diversity. Transformations such as rotation, flipping, zooming, and translation are applied
randomly to generate new training samples from existing ones. These transformations simulate
real-world variations in patient positioning or imaging angles, allowing the model to become more
robust to spatial distortions. Additionally, brightness and contrast jittering is applied to help the
model adapt to illumination variations. Care is taken to ensure that augmentation does not alter the
fundamental anatomical structure of the brain, preserving the biological integrity of the data.

Through augmentation, the effective training dataset size increases several-fold, significantly
improving model generalization and reducing overfitting tendencies that could otherwise arise
from limited original data.

Impact of Preprocessing on Model Performance

The preprocessing pipeline significantly enhances the reliability and interpretability of the final
model. Through standardization, normalization, and augmentation, the system becomes resilient
to scanner variations and capable of focusing on clinically meaningful patterns. During preliminary
experiments, it was observed that models trained on raw, unprocessed images exhibited erratic
behaviour and unstable learning curves, whereas the same architectures trained on pre-processed
data achieved faster convergence, higher validation accuracy, and improved generalization across
all diagnostic [Link], the inclusion of skull stripping and intensity normalization
proved particularly valuable in emphasizing Alzheimer-related brain regions, as verified through
Grad-CAM visualizations. This demonstrates that the preprocessing steps not only improve
numerical performance but also enhance the biological relevance of learned features.

38
 CHAPTER 4.
RESULTS ANALYSIS AND VALIDATION

4.1 Overview of Experimental Outcomes

The effectiveness of any deep learning-based diagnostic system is ultimately determined by the
quality of its results and the reliability of its validation process. After extensive experimentation
and careful model training, the proposed 2D Hybrid Neural Network for Alzheimer’s Disease
Detection was subjected to a comprehensive evaluation framework. The primary purpose of this
phase was not merely to achieve high numerical accuracy but to establish clinical validity,
interpretability, and consistency across multiple evaluation parameters. This section presents an
in-depth analysis of the obtained results, discussing the performance metrics, model behavior,
validation strategy, and implications for real-world applicability.

The experimental setup was designed to evaluate the hybrid 2D convolutional architecture on the
pre-processed MRI dataset discussed in earlier sections. The dataset was partitioned into three
subsets: approximately 70% for training, 15% for validation, and 15% for testing. This division
ensured that the model’s evaluation remained independent and unbiased. During the training
phase, the model was trained for multiple epochs using an Adam optimizer with a learning rate
of 0.0001, employing categorical cross-entropy as the loss function. The training and validation
losses were continuously monitored to detect convergence and prevent overfitting. Early stopping
mechanisms were employed to terminate training once the validation performance plateaued,
ensuring computational efficiency and model generalization.

The model’s output consisted of four probabilistic classifications — Non-Demented, Very Mild
Demented, Mild Demented, and Moderate Demented — corresponding to the diagnostic stages
present in the dataset. The final trained model demonstrated high accuracy and robust convergence,
validating the architectural design and feature preprocessing strategy. On the test dataset, the
model achieved a mean classification accuracy exceeding 94%, accompanied by strong precision,

39
recall, and F1-scores across all four categories. The Receiver Operating Characteristic (ROC)
curves for each class displayed consistent high areas under the curve (AUC values between 0.95
and 0.98), indicating exceptional discriminative capability. These outcomes suggest that the model
successfully captures the neuroanatomical variations associated with different stages of
Alzheimer’s disease.

Analysis of Model Convergence and Training Behaviour

The training and validation curves revealed an important aspect of the model’s learning dynamics.
During the initial epochs, both the training and validation losses decreased rapidly, reflecting the
network’s ability to extract meaningful patterns from the MRI data. As training progressed, the
loss values gradually stabilized, indicating convergence to an optimal point. Unlike many deep
models that suffer from overfitting when trained on limited medical data, the proposed hybrid
model demonstrated stable behaviour due to effective regularization mechanisms such as dropout
and batch normalization. The inclusion of data augmentation further contributed to this stability
by providing diverse and representative input samples, allowing the model to generalize across
unseen data variations.

The smooth convergence pattern is a strong indicator of a well-tuned model architecture and
confirms that the preprocessing steps successfully removed redundant variability while preserving
essential diagnostic information. The close alignment between training and validation curves also
suggests that the model does not rely on memorization but learns robust feature representations
that are transferrable across different patient cases.

Performance Evaluation Metrics

In medical diagnostics, relying solely on accuracy can be misleading, particularly when dealing
with imbalanced datasets. Therefore, multiple evaluation metrics were utilized to assess the
model’s performance comprehensively, each representing a unique dimension of diagnostic
[Link] model achieved an overall accuracy of 94–96% on the test dataset, which indicates
that the majority of MRI scans were correctly classified into their respective categories. However,
beyond accuracy, more insightful metrics provide a deeper understanding:

40
  Precision values across all classes were above 93%, confirming that the system generates
minimal false positives. In medical terms, this means that most cases identified as
Alzheimer’s truly correspond to the disease, reducing the likelihood of unnecessary
medical interventions.
 Recall (Sensitivity) averaged around 95%, which implies that the model is highly sensitive
to identifying Alzheimer’s patients and rarely misses true cases. High recall is particularly
vital for early detection, as failing to identify diseased individuals can delay treatment and
reduce patient outcomes.
 The F1-Score, the harmonic mean of precision and recall, remained consistently high
(between 0.94 and 0.95), validating the model’s balance between sensitivity and
specificity.
 The ROC-AUC metric further reinforced these findings, with values exceeding 0.95 for
all four diagnostic categories. This high AUC value reflects the system’s excellent ability
to distinguish between Alzheimer’s stages, even under varying threshold conditions.

The distribution of these performance metrics reveals a model that is not only accurate but also
clinically trustworthy. By achieving consistently strong results across multiple statistical
dimensions, the hybrid 2D CNN model demonstrates a capacity for reliable and repeatable
diagnosis in real-world scenarios.

Validation through Explainable AI

While numerical results provide strong evidence of model performance, clinical applications
demand interpretability. To validate that the model’s predictions are grounded in biologically
meaningful regions, explainable AI (XAI) techniques were employed. The Gradient-weighted
Class Activation Mapping (Grad-CAM) method was particularly used to visualize which areas of
MRI images most influenced the model’s decisions.

The Grad-CAM heatmaps consistently highlighted the hippocampus, medial temporal lobe, and
cortical regions, all of which are well-established biomarkers of Alzheimer’s pathology. These
visualizations confirm that the model’s learning process aligns with medical understanding — it
focuses on anatomically relevant brain structures rather than irrelevant or random image regions.

41
This interpretability aspect is essential for establishing clinical trust. By showing that the neural
network bases its decisions on valid anatomical evidence, the system moves beyond being a “black
box” to a transparent and supportive diagnostic tool. Moreover, these explainable insights can
assist radiologists in cross-verifying AI results, fostering a human–AI collaborative framework for
Alzheimer’s diagnosis.

Limitations and Observations

While the model demonstrates excellent accuracy and interpretability, certain limitations were
observed. The primary challenge lies in distinguishing between very mild and mild Alzheimer’s
stages, where structural changes are minute and often indistinguishable in isolated 2D slices. This
limitation stems from the inherent biological overlap between adjacent disease stages rather than
algorithmic shortcomings.

Another challenge relates to dataset diversity. Although the chosen dataset provides clean and
well-labelled MRI scans, it lacks demographic diversity and may not fully represent variations
across ethnicities, scanner models, or acquisition conditions. Future work should include multi-
site datasets and multimodal data (e.g., PET or fMRI) to enhance model generalization. Despite
these limitations, the overall performance metrics affirm that the proposed approach effectively
addresses the key research problem — achieving high diagnostic accuracy and interpretability
while maintaining low computational requirements.

4.2 System Architecture and Workflow

The architecture and workflow of an artificial intelligence system determine not only its
computational efficiency but also its capacity to extract, learn, and generalize meaningful
patterns from data. In the context of Alzheimer’s disease detection using MRI imaging, the
architecture must effectively capture the subtle neuroanatomical variations associated with
progressive brain atrophy while remaining computationally lightweight and clinically
interpretable. The proposed 2D Hybrid Neural Network (2D-HNN) is an advanced deep learning
framework specifically designed to achieve this balance. It combines the spatial sensitivity of
convolutional networks with the representational power of dense layers and integrates attention-

42
driven aggregation to capture inter-slice dependencies. The system follows a structured
workflow beginning from data acquisition and preprocessing to feature extraction, model
training, validation, and interpretability analysis
Conceptual Overview of System Design
The overall architecture of the proposed system follows a layered modular design, where each
module performs a specific task in the Alzheimer’s detection pipeline. The structure is divided
into five major stages:
1. Input and Preprocessing Layer – Handles data ingestion, normalization, and
augmentation.
2. Feature Extraction Layer – Employs convolutional filters to capture spatial and textural
features from 2D MRI slices.
3. Hybrid Learning Layer – Combines convolutional outputs with dense fully connected
layers to perform nonlinear mapping and classification.
4. Output Layer – Produces probabilistic predictions across four diagnostic categories.
5. Interpretability and Validation Layer – Generates Grad-CAM visualizations and
evaluates model reliability using performance metrics.
This modular structure allows for scalability and adaptability — enabling further integration with
multimodal imaging data (e.g., PET, fMRI) or clinical biomarkers in future iterations.

Model Training Workflow

The training workflow defines how the model learns optimal parameters through iterative
exposure to data. It follows a supervised learning paradigm, where the model compares its
predicted outputs against known labels and adjusts its weights accordingly to minimize error.
During training, the model employs the Adam optimizer, chosen for its adaptive learning rate
capabilities and efficient handling of sparse gradients. The learning rate was initially set to
0.0001 after experimental tuning. The categorical cross-entropy loss function measures the
divergence between predicted probabilities and true class [Link] training process is
structured into epochs, where each epoch represents a complete pass through the training dataset.
Empirical analysis revealed that the model converged optimally around 25–30 epochs, achieving
a plateau in validation accuracy beyond that point. Early stopping was used to automatically halt
training once the validation loss ceased improving, preventing overfitting and conserving

43
computational resources.
Throughout the training process, the model parameters — weights and biases — are updated
using backpropagation, where gradients of the loss function with respect to each parameter are
computed and used to perform incremental adjustments. The integration of batch normalization,
dropout, and augmented datasets ensures that the optimization process remains stable and
resistant to local minima.
On average, the full model training required approximately 2.5 to 3 hours on an NVIDIA RTX
3060 GPU using a dataset of around 5,000 MRI slices. The training accuracy steadily increased
from 78% in early epochs to over 95% by the final epoch, demonstrating strong learning
efficiency and convergence stability.
Performance Statistics and Observations
Quantitative assessment of the architecture reveals its strength in balancing speed, accuracy, and
interpretability.
Key performance statistics recorded during model training include:

 Training Time: 2.5–3 hours on RTX 3060 GPU

 Training Accuracy: ~95.8%
 Validation Accuracy: ~94.6%
 Test Accuracy: ~94.1%
 Average F1-Score: 0.94
 Mean ROC-AUC: 0.96
 Average Inference Time per Image: 45–60 milliseconds

These results confirm that the architecture efficiently processes high-dimensional MRI data
while maintaining low latency — a crucial requirement for real-time or near-real-time diagnostic
applications.
The relatively small inference time makes the model suitable for integration into clinical imaging
workflows where rapid analysis is necessary. Furthermore, the strong alignment between
validation and test metrics indicates consistent performance and minimal overfitting.

44
 CHAPTER 5.
FUTURE PROSPECTS AND CONCLUSION

5.1. Summary of Achievements

The journey from conceptualization to implementation of the proposed Alzheimer Detection via
2D Hybrid Neural Networks represents a significant stride toward the integration of artificial
intelligence and medical imaging for early neurodegenerative disease diagnosis. This study has
demonstrated that deep learning models, when properly structured, trained, and validated, can
extract subtle patterns from MRI data that often elude manual observation. Through the application
of advanced preprocessing techniques, optimized network architectures, and interpretable AI
methodologies, the project successfully establishes an intelligent diagnostic framework that is both
accurate and computationally efficient.

The research achieved multiple milestones across conceptual, methodological, and experimental
stages. At its core, the proposed system is built upon a 2D Hybrid Neural Network (HNN)
architecture — a design that merges convolutional and dense neural networks to capture both
spatial and semantic representations of brain structures from MRI scans. By systematically
integrating this architecture within a carefully structured workflow, the system achieved robust
classification accuracy across multiple Alzheimer’s stages — Non-Demented, Very Mild
Demented, Mild Demented, and Moderate Demented.

The model attained an overall test accuracy of approximately 94–96%, with high precision,
recall, and F1-scores (≈0.94) across all categories. Furthermore, the ROC-AUC values exceeding
0.95 confirm that the system reliably distinguishes between different disease stages, even under
varying classification [Link] addition to numerical excellence, the model’s interpretability
was validated using Gradient-weighted Class Activation Mapping (Grad-CAM), which
highlighted biologically relevant regions such as the hippocampus, temporal lobes, and cortical
thinning areas — precisely those regions known to deteriorate during Alzheimer’s progression.
This alignment between computational focus and medical knowledge demonstrates that the model
is not a “black box” but an explainable, trustworthy diagnostic assistant.

45
From a technical standpoint, the project achieved significant computational efficiency. By
adopting a 2D rather than a 3D convolutional approach, the training time was reduced from an
estimated 12–15 hours (for volumetric CNNs) to less than 3 hours on a standard NVIDIA RTX
3060 GPU, with an average inference time of 50 milliseconds per MRI slice. This efficiency
ensures that the system remains deployable in modest research laboratories, small hospitals, and
resource-limited clinical environments, thereby democratizing access to AI-driven diagnostic
tools.

Limitations of the Current Study

Despite its promising results, the study acknowledges certain limitations that provide avenues for
[Link] first limitation pertains to dataset diversity. The dataset used, while clean and
well-labeled, was limited in size and geographic diversity. Since MRI acquisition protocols vary
across hospitals and scanner manufacturers, models trained on a single dataset may experience
performance degradation when applied to different data sources. Future research must address this
limitation by incorporating multi-centre datasets with varied imaging parameters and
demographics to enhance generalization.

Secondly, the current implementation relies solely on structural MRI data. However, Alzheimer’s
disease involves multiple pathophysiological changes, some of which may not be detectable
through structure alone. Incorporating multimodal imaging data — such as PET (Positron
Emission Tomography) or functional MRI (fMRI) — could provide deeper insights into metabolic
and functional alterations, improving predictive accuracy and clinical interpretability. Thirdly,
while the current architecture achieves interpretability through Grad-CAM visualization, the field
of explainable AI continues to evolve. Future enhancements could include attention-based
mechanisms or feature attribution methods like SHAP (Shapley Additive Explanations) to provide
more granular interpretive insights.

Finally, ethical and regulatory challenges remain. Deployment in clinical environments would
require compliance with medical data protection laws such as HIPAA and GDPR, rigorous clinical
trials, and real-world testing under professional supervision. These steps are necessary to transition
the system from an academic prototype to a certified medical-grade diagnostic aid

46
5.2 Future Prospects

The future of AI in neurodegenerative disease diagnosis is exceptionally promising, and the

present research lays the groundwork for multiple extensions and innovations.

One promising direction involves the integration of 3D contextual understanding into the current
2D framework. By combining the efficiency of 2D convolutional processing with 3D spatial
correlation modelling, future architectures could achieve higher diagnostic fidelity without
drastically increasing computational cost. Hybrid 2D–3D architectures are already emerging in
recent medical imaging research and could serve as the next logical step for enhancing model
precision. Another major prospect lies in the development of longitudinal analysis models —
systems that track changes in a patient’s MRI scans over time to monitor disease progression. Such
models could enable predictive analytics capable of forecasting Alzheimer’s trajectory, thereby
supporting treatment planning and personalized intervention strategies.

Additionally, integrating genomic and clinical data (such as age, family history, cognitive test
results, and APOE gene information) could further enrich the model’s predictive capability. A
multimodal fusion of imaging and non-imaging data could transform the model from a diagnostic
tool into a comprehensive decision-support system, capable of holistic patient assessment. From a
deployment perspective, the future also envisions cloud-based and edge AI solutions. With
advancements in medical IoT and federated learning, the model could be implemented in
distributed healthcare networks where MRI data remains localized for privacy, while shared model
parameters continuously update through federated training. This decentralized approach enhances
both data privacy and global adaptability.

Concluding Remarks

In conclusion, this project demonstrates the immense potential of AI-driven neuroimaging analysis
in transforming how Alzheimer’s disease is detected, analysed, and understood. The proposed 2D
Hybrid Neural Network framework successfully combines computational efficiency, diagnostic

47
accuracy, and medical interpretability into a unified system capable of supporting clinical decision-
making.

By achieving high classification performance, aligning computational focus with biological

relevance, and operating efficiently on standard hardware, the project fulfills its central objective:
to create a scalable, reliable, and transparent diagnostic model for Alzheimer’s [Link]
importantly, beyond its technical success, this research contributes to a larger paradigm shift —
from manual, subjective diagnostic methods toward data-driven precision medicine. The use of
interpretable deep learning systems can empower clinicians with quantitative evidence, reduce
diagnostic disparities, and accelerate early intervention strategies — ultimately improving patient
outcomes and quality of life.

While the journey toward clinical adoption requires further validation, ethical frameworks, and
regulatory compliance, this study establishes a solid foundation for future innovation. The 2D
Hybrid Neural Network model stands not merely as a research prototype but as a tangible step
toward AI-assisted healthcare democratization — where intelligent systems enhance human
expertise to detect diseases earlier, faster, and more accurately than ever before.

48
 REFERENCES

[1] M. A. Joy, S. Nasrin, A. Siddiqua, and D. M. Farid, “ViTAD: Leveraging modified vision
transformer for Alzheimer’s disease multi-stage classification from brain MRI scans,” Preprint
submitted to Elsevier, Jan. 2025.

[2] S. E. Sorour, A. A. Abd El-Mageed, K. M. Albarrak, A. K. Alnaim, A. A. Vafa, and E. El-

Shafiey, “Classification of Alzheimer’s disease using MRI data based on deep learning
techniques,” J. King Saud Univ. - Comput. Inf. Sci., vol. 36, no. 2, pp. 101940, 2024.

[3] Z. Zhao, S. Wu, and J. Chen, “Vision transformer–equipped convolutional neural networks
for Alzheimer’s disease diagnosis using 3D MRI,” PMC, Dec. 2024.

[4] S. Alp, M. Koyuncu, and R. Meier, “Joint transformer architecture in brain 3D MRI
classification for Alzheimer’s disease,” Nat. Sci. Rep., vol. 14, pp. 5578, 2024.

[5] F. Haja Mohideen, R. Abdulazeez, D. Morrow, Y. Sheykhahmad, N. Rauf, and A. Waris,

“Four-way classification of Alzheimer’s disease using deep learning with MRI data,” PMC,
Feb. 2023.

[6] J. C. Bortty, M. Rahman, and S. Islam, “A novel diagnostic framework with an optimized
ensemble of vision transformers and convolutional neural networks for enhanced Alzheimer’s
disease detection,” PMC, Mar. 2025.

[7] F. Isensee, M. Schell, I. Tursunova, G. Brugnara, D. Bonekamp, U. Neuberger, A. Wick, H.

P. Schlemmer, S. Heiland, W. Wick, M. Bendszus, K. H. Maier-Hein, and P. Kickingereder,
“Automated brain extraction of multisequence MRI using artificial neural networks,” Hum.
Brain Mapp., vol. 40, no. 17, pp. 4952–4964, 2019.

[8] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M.

Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth
16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv:2010.11929,
2020.

49
[9] Z. Liu, H. Mao, C. Y. Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A ConvNet for the
2020s,” in Proc. IEEE/CVF Conf. Comput. Vision Pattern Recognit., 2022, pp. 11976–11986.

[10] D. S. Marcus, T. H. Wang, J. Parker, J. G. Csernansky, J. C. Morris, and R. L. Buckner,

“Open access series of imaging studies (OASIS): Cross-sectional MRI data in young, middle
aged, nondemented, and demented older adults,” J. Cogn. Neurosci., vol. 19, no. 9, pp. 1498–
1507, 2007.

[11] R. C. Petersen, P. S. Aisen, L. A. Beckett, M. C. Donohue, A. Gamst, D. J. Harvey, C. R.

Jack, W. J. Jagust, L. M. Shaw, A. W. Toga, J. Q. Trojanowski, and M. W. Weiner, “Alzheimer’s
Disease Neuroimaging Initiative (ADNI): Clinical characterization,” Neurology, vol. 74, no. 3,
pp. 201–209, 2010.

[12] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in
Proc. IEEE Conf. Comput. Vision Pattern Recognit., 2016, pp. 770–778.

[13] G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected
convolutional networks,” in Proc. IEEE Conf. Comput. Vision Pattern Recognit., 2017, pp.
4700–4708.

[14] K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image
recognition,” arXiv preprint arXiv:1409.1556, 2014.

[15] M. Tan and Q. Le, “EfficientNet: Rethinking model scaling for convolutional neural
networks,” in Proc. Int. Conf. Mach. Learn., 2019, pp. 6105–6114.

[16] J. Islam and Y. Zhang, “Brain MRI analysis for Alzheimer’s disease diagnosis using an
ensemble system of deep convolutional neural networks,” Brain Inform., vol. 5, pp. 1–14, 2018.

[17] S. Sarraf and G. Tofighi, “Deep learning-based pipeline to recognize Alzheimer’s disease
using MRI data,” in Proc. Future Technol. Conf., IEEE, 2016, pp. 816–820.

50
[18] Malik, H. Ahmed, S. Farooq, T. Sadiq, F. Amin, A. Hussain, M. Habib, A. Islam, and S.
Khan, “Deep learning for Alzheimer’s disease prediction: A comprehensive review,” PMC, Jun.
2024.

[19] T. Tuncer, S. Dogan, and A. Subasi, “FiboNeXt: Investigations for Alzheimer’s disease
detection using fibonacci sequence-based ConvNeXt architecture,” Biomed. Signal Process.
Control, vol. 90, pp. 105313, 2025.

[20] L. Pei, S. Bakas, A. Vosoughi, S. Reza, D. Davatzikos, and C. Iftekharuddin, “A general

skull stripping of multiparametric brain MRIs using 3D convolutional neural network,” Nat. Sci.
Rep., vol. 12, pp. 10826, 2022.

51