PROJECT REPORT

On
“Deep Learning-Based Classification of Alzheimer's Disease
Stages Using MRI Scans and VGG16 Architecture”
A Mini Project
Submitted in Partial Fulfillment of the Requirements for the
Degree of
[Link]. Data Science and Statistics
(III Semester)
Submitted By
Shivani Gusain
Aditya Gusain
Subrat Kothiyal

Under the Guidance of

Dr. Ankit Tomar
(Associate Professor)

DEPARTMENT OF MATHEMATICS
GRAPHIC ERA HILL UNIVERSITY
2024-2025
 ABSTRACT

Alzheimer's disease is a degenerative neurological condition and early detection plays a

vital role in effective treatment. This study leverages the VGG16 Convolutional Neural
Network (CNN) framework to classify different stages of Alzheimer's disease using a
dataset comprising approximately 12,000 MRI images. The preprocessing pipeline
includes resizing, normalization, and data augmentation to improve the model's ability
to generalize. The network is fine-tuned with pre-trained ImageNet weights and
integrates batch normalization and dropout layers to minimize the risk of overfitting.
To assess performance, evaluation metrics such as accuracy, precision, and recall are
utilized, along with loss curves for detailed analysis. Optimization techniques,
including early stopping and learning rate adjustments, are applied to enhance the
training process. This research demonstrates the effectiveness of CNNs and transfer
learning in building AI-powered diagnostic systems for healthcare applications.

Keywords- Alzheimer's disease, Deep-Learning, MRI, VGG16, High Classification Accuracy,

CNN.
 INTRODUCTION

1. Introduction

Alzheimer's Disease (AD) is a brain disorder that gradually affects memory, thinking,
and behavior. It is the most common cause of dementia, responsible for 60-80% of
cases (Alzheimer's Association, 2021). Early diagnosis is important for managing
symptoms and slowing the progression of the disease. However, traditional methods
like clinical assessments and memory tests can be time-consuming and may lack
accuracy (Jack et al., 2018).

Recent advances in deep learning have transformed medical image analysis, enabling
highly accurate and automated diagnoses. Convolutional Neural Networks (CNNs) are
widely used for image classification, especially in medical imaging (Litjens et al.,
2017). One such model, VGG16, introduced by Simonyan and Zisserman (2014), has
shown excellent performance in recognizing patterns and extracting features due to its
deep architecture.

1.2 Importance of Early Detection

Early detection is vital for timely interventions, improving patient quality of life
through therapies that slow cognitive decline (Dubois et al., 2016). Imaging methods
like MRI and PET detect brain changes but are prone to errors in manual analysis,
necessitating automated approaches (Jack et al., 2010). Deep learning models can
identify early patterns indicative of AD, enabling faster, more accurate diagnoses and
supporting clinical trials for new treatments (Petersen et al., 2014).

1.3 Transfer Learning in Healthcare

Transfer learning is a machine learning approach that leverages knowledge gained from
one domain to improve performance in another domain. In the context of medical
imaging, pre-trained models, such as VGG16, can be fine-tuned for specific tasks,
significantly reducing training time and computational resources (Pan & Yang, 2010).
Transfer learning has proven particularly effective when labelled data is scarce, which
is often the case in healthcare applications (Shin et al., 2016).

Studies have demonstrated the efficacy of transfer learning in diagnosing various

medical conditions, including diabetic retinopathy, breast cancer, and lung diseases
(Esteva et al., 2017; Rajpurkar et al., 2017). For Alzheimer's Disease, transfer learning
models can extract high-level features from MRI scans, enabling accurate classification
into different stages of dementia (Sarraf & Tofighi, 2016). Such methods outperform
traditional machine learning algorithms by leveraging pre-trained weights optimized for
general image recognition tasks (Gulshan et al., 2016).

1.4 Proposed Approach

This project focuses on leveraging the VGG16 model for classifying Alzheimer's
Disease using MRI scans. The approach involves transfer learning, where pre-trained
weights are fine-tuned to adapt the model for Alzheimer's classification. Transfer
learning reduces training time and improves accuracy, especially when dealing with
limited datasets (Pan & Yang, 2010).

In this study, we evaluate the performance of VGG16 in distinguishing between four

categories: Non-Demented, Very Mild Demented, Mild Demented, and Moderate
Demented. Performance metrics such as accuracy, precision, recall, and F1-score are
used to assess the model's effectiveness. This work aims to contribute to the ongoing
research in AI-based medical diagnostics and provide a scalable solution for early
Alzheimer's detection.
 2. MOTIVATION AND CONTRIBUTION

2.1 Motivation:

Alzheimer's Disease (AD) remains one of the leading causes of dementia, impacting
millions of individuals worldwide. Early detection of AD is essential for effective
treatment and management, as it provides opportunities for timely interventions that
may slow disease progression and improve patient outcomes. Conventional diagnostic
approaches, such as clinical evaluations and neuroimaging, are often labor-intensive,
subjective, and prone to variability. With the increasing volume of medical data, there
is an urgent need for automated and accurate diagnostic tools.

Deep learning has emerged as a transformative technology in medical imaging, offering

high accuracy and efficiency. Convolutional Neural Networks (CNNs), particularly
pre-trained models like VGG16, provide powerful feature extraction capabilities,
making them well-suited for Alzheimer's classification tasks. Leveraging transfer
learning allows us to address the challenges posed by limited labeled data and complex
image patterns in medical datasets. This project aims to bridge the gap between cutting-
edge AI techniques and clinical diagnostics by focusing on Alzheimer's Disease
classification using VGG16.

2.2 Related Work:

Alzheimer’s disease detection has greatly benefited from advancements in machine

learning and deep learning techniques. These methods leverage neuroimaging datasets
like MRI and PET scans for diagnosis. Over the years, researchers have developed
various models, including CNNs, hybrid techniques, and attention mechanisms, to
improve the accuracy of detection and classification.

Below is a summary of state-of-the-art studies, organized in ascending order by

publication year:
2.3 Contribution:

This research makes the following key contributions:

1. Application of Transfer Learning with VGG16: Leverages the pre-trained

VGG16 model for Alzheimer's Disease classification, using its powerful feature
extraction capabilities to identify different stages of dementia. This reduces the
need for large labeled datasets while achieving high performance in MRI image
analysis. The model’s deep architecture enhances sensitivity to subtle brain
changes, improving diagnostic accuracy.
2. Development of an Automated Diagnostic Tool: Proposes an efficient,
scalable pipeline for MRI-based Alzheimer's classification, automating the
diagnostic process and reducing manual effort. The tool ensures consistent
results and improves accuracy, making it suitable for integration into clinical
settings. Its scalable design allows for easy adaptation to larger datasets and
broader applications.

3. Fine-Tuning for Enhanced Performance: Fine-tunes the VGG16 model using

transfer learning techniques to optimize it for Alzheimer’s detection. Data
augmentation and advanced optimization algorithms like Adam help the model
generalize better on limited datasets. This fine-tuning boosts the model's
sensitivity and specificity, ensuring it performs well even on smaller, domain-
specific datasets.

4. Comprehensive Evaluation Metrics: Evaluates the model using accuracy,

precision, recall, F1-score, and AUC, ensuring robustness and reliability in
detecting Alzheimer’s. These metrics balance performance, minimizing false
positives and negatives while maximizing diagnostic accuracy. The evaluation
also guarantees the model’s effectiveness in differentiating Alzheimer’s stages.

5. Practical Implementation Framework: Provides a step-by-step framework for

implementing transfer learning in Alzheimer’s detection, which can be adapted
to other medical imaging tasks. This framework covers data preprocessing,
model training, and deployment, with a focus on transparency and
interpretability for clinical adoption. It demonstrates how AI can be extended to
a range of healthcare applications.
 3. RESEARCH METHODOLOGY

3.1 Dataset:

The dataset used in this study was obtained from an open online dataset library known
as Kaggle. This dataset contains almost 12,000 images distributed across train and test
folders. Each folder consists of MRI brain scan images classified into four categories:
Mild Demented, Moderate Demented, Non-Demented, and Very Mild Demented. The
training dataset is then split into 80% training and 20% validation datasets. The dataset
is structured to support supervised learning tasks, making it suitable for multi-class
classification problems. Each class represents a different stage of Alzheimer's disease,
enabling the model to distinguish between varying levels of cognitive impairment. The
dataset includes separate folders for training and testing, simplifying model evaluation
and performance monitoring. Images were resized to 224 x 224 pixels to ensure
compatibility with pre-trained deep learning architectures. Additionally, data
augmentation techniques such as rotation, flipping, zooming, and shifting were applied
to increase variability, prevent overfitting, and enhance model generalization. This
dataset serves as a comprehensive resource for the development and validation of
convolutional neural network (CNN)-based models for Alzheimer's classification tasks.

Non-demented Very Mild demented Mild demented Moderate demented

Non-demented Very Mild demented Mild demented Moderate demented

Fig. 1. Dataset after Pre-processing

3.2 METHODOLOGY

Fig. 2. Proposed Methodology

The methodology of this study involves several key stages to prepare and analyze the
dataset for Alzheimer's disease classification. The flowchart below illustrates the steps
involved:

1. Raw Data: The dataset is obtained from an open online dataset library known
as Kaggle. This dataset contains almost 12,000 MRI brain scan images
distributed across train and test folders.
2. Preprocessing: Preprocessing techniques are applied to prepare the raw data for
analysis. This includes resizing images to 224x224 pixels to ensure
compatibility with pre-trained deep learning architectures, normalization, and
data augmentation (e.g., rotation, flipping, zooming, and shifting). These steps
enhance model generalization and prevent overfitting.
3. Processed Data: The processed data is split into 80% training and 20%
validation datasets. This structure supports supervised learning tasks and
simplifies model evaluation and performance monitoring.
4. Feature Extraction: In this stage, features relevant to Alzheimer's disease
classification are extracted from the processed data using deep learning models.
The VGG16 model, initialized with pre-trained weights from ImageNet, is used
for feature extraction.
5. Feature Selection: Additional layers incorporating batch normalization and
dropout are added to the VGG16 model to fine-tune it for the specific task of
classifying Alzheimer's disease stages. This step helps in selecting the most
relevant features and further prevents overfitting.
6. Modeling: The final stage involves training the model using the extracted and
selected features. The performance of the model is evaluated using metrics such
as accuracy, precision, recall, and loss curves. Early stopping and learning rate
scheduling are implemented to optimize the training process.
3.3 MODEL ARCHITECTURE:

Convolutional Neural Networks (CNNs)

Convolutional Neural Networks (CNNs) are a specialized class of deep learning models
designed to analyze visual data. The process begins by passing an image through
convolutional layers, where filters are applied to detect basic patterns like edges and
textures. The output of these layers is then passed through a ReLU activation function,
introducing non-linearity and enabling the model to learn complex features. Pooling
layers follow, reducing the size of the feature maps while retaining important
information, which helps in reducing computational cost and avoiding overfitting.

Afterward, the data is flattened into a 1D vector and passed through fully connected
layers to make predictions based on the learned features. The final output layer uses a
softmax activation function to assign probabilities to each class, determining the most
likely category for the image. Pre-trained CNN models like VGG16 and ResNet
facilitate transfer learning, allowing these models to be adapted for specific tasks with
smaller datasets. This makes CNNs highly efficient for a variety of applications,
including medical image analysis and object detection.

In this study, CNNs were utilized to classify Alzheimer's disease stages from MRI
images, leveraging their ability to extract high-level features critical for identifying
patterns in medical data.

Fig. 3. CNN Architecture

VGG16

VGG16 is a widely used convolutional neural network (CNN) architecture introduced

by the Visual Geometry Group at the University of Oxford. It was developed as part of
the ILSVRC-2014 (ImageNet Large Scale Visual Recognition Challenge), where it
demonstrated exceptional performance in image classification tasks. The architecture
consists of 16 layers with learnable parameters, including 13 convolutional layers and
3 fully connected layers.

VGG16 employs 3x3 convolutional filters with a stride of 1, enabling it to capture

intricate patterns and features within images. It also utilizes 2x2 max-pooling layers for
down-sampling, which reduces spatial dimensions while preserving important features.
The model incorporates ReLU (Rectified Linear Unit) as its activation function to
introduce non-linearity and improve learning capabilities.

The final layers of VGG16 include fully connected layers followed by a softmax layer
for multi-class classification tasks. Its input size requirement is 224x224x3, making it
suitable for RGB image datasets. Pre-trained weights, particularly those trained on the
ImageNet dataset, allow VGG16 to be utilized effectively for transfer learning
applications.

In this study, the VGG16 architecture was considered due to its deep hierarchical
structure, which enables efficient feature extraction, and its proven ability to handle
complex image classification tasks. Despite having approximately 138 million
parameters, VGG16 remains a benchmark in computer vision tasks and is frequently
adopted for medical imaging research, including the classification of Alzheimer's
disease using MRI datasets.

Fig.4 VGG16 Architecture

3.4 Training Process:

1. Loss Function: Categorical Crossentropy.

2. Optimizer: Adam optimizer with a learning rate scheduler (initial learning rate =
0.0001).
3. Metrics: Accuracy, Precision, Recall, F1-score, and ROC-AUC for detailed
performance evaluation.
4. Dataset Split: 90% training, 10% validation, and separate testing to ensure
comprehensive evaluation.
 5. Early Stopping: Implemented to monitor validation loss and avoid overfitting.
6. Learning Rate Scheduler: Reduces learning rate when validation loss plateaus.

3.5 Tools and Frameworks:

1. Programming Language: Python (Version 3.9.7).

2. Libraries and Frameworks:
 TensorFlow and Keras for model implementation and training.
 NumPy and Pandas for data manipulation.
 Matplotlib and Seaborn for data visualization and performance plotting.
 Scikit-learn for evaluation metrics and additional preprocessing.
3. Development Environment:
 Jupyter Notebooks for coding, training, and visualization.
 GPU support enabled for faster training.
 4. RESULTS AND DISCUSSION

4.1 VGG16 Model Results:

The VGG16 Model was trained using 9216 MRI images, validated on 1024 MRI
images and tested on 1280 MRI images. It used a batch size of 22 images to train itself.
The model has runned for 32 epochs. The model provides an accuracy of about 88% in
the training data, 79% on validation data and about 75% in testing data. Furthermore, in
the confusion matrix, the number on the diagonal indicates the number of times the
samples were classified correctly, the numbers that are not on the diagnol show how
many times the samples were classified incorrectly.

Based on the classification report, it is evident that the model performs best on the
"Moderate Impairment" class, with a precision of 92% and a recall of 100%, resulting
in an F1-score of 96%. The model's performance is slightly lower for the "Very Mild
Impairment" class, where the F1-score is 78%, indicating a potential area for
improvement. The overall F1-scores for the other classes ("Mild Impairment" and "No
Impairment") are also strong, suggesting that the model is robust across the different
classes. However, further tuning could help improve the results for the least well-
performing class.

These results reflect the model’s ability to classify the different stages of Alzheimer's
disease accurately, though it may still benefit from strategies such as additional data
augmentation, fine-tuning of hyperparameters, or the use of more advanced
architectures to further optimize its [Link] graphs, Confusion matrix and
classification report are as follows:
 Fig. 5. Model Loss in VGG16

Fig. 6. Model Accuracy in VGG16

Fig. [Link] Matrix for VGG16

Fig. 8. Classification Report

 5. CONCLUSION

This study demonstrates the effectiveness of transfer learning in detecting the stages of
Alzheimer's disease from medical images. A VGG16 model was employed for
classifying MRI images, with the model being trained on a dataset of 9216 images,
validated on 1024 images, and tested on 1280 images. The model was trained over 32
epochs with a batch size of 22. The results indicate that the model performed well on
the training data with an accuracy of 85%. However, its performance on the validation
and test datasets was somewhat lower, achieving accuracies of 79% and 85%,
respectively. The confusion matrix analysis revealed that the model was able to
correctly classify a significant number of samples, with a few misclassifications
observed, particularly in the off-diagonal elements. These misclassifications provide
valuable insights for potential areas of improvement in the model's performance.

6. REFERENCES

 Alzheimer's Association. (2021). 2021 Alzheimer's Disease Facts and Figures.

Alzheimer's & Dementia, 17(3), 327-406.
 Jack, C. R., et al. (2018). NIA-AA Research Framework: Toward a biological
definition of Alzheimer's disease. Alzheimer's & Dementia, 14(4), 535-562.
 Litjens, G., et al. (2017). A survey on deep learning in medical image analysis.
Medical Image Analysis, 42, 60-88.
 Simonyan, K., & Zisserman, A. (2014). Very Deep Convolutional Networks for
Large-Scale Image Recognition. arXiv preprint arXiv:1409.1556.
 Frisoni, G. B., et al. (2010). The clinical use of structural MRI in Alzheimer
disease. Nature Reviews Neurology, 6(2), 67-77.
 Sarraf, S., & Tofighi, G. (2016). DeepAD: Alzheimer's Disease Classification
via Deep Convolutional Neural Networks using MRI and fMRI. bioRxiv,
070441.
 Pan, S. J., & Yang, Q. (2010). A Survey on Transfer Learning. IEEE
Transactions on Knowledge and Data Engineering, 22(10), 1345-1359.
 Dubois, B., et al. (2016). Preclinical Alzheimer's disease: Definition, natural
history, and diagnostic criteria. Alzheimer's & Dementia, 12(3), 292-323.
 Jack, C. R., et al. (2010). Brain atrophy rates predict subsequent clinical
conversion in normal elderly and amnestic MCI. Neurology, 74(7), 687-693.
 Liu, S., et al. (2014). Multimodal neuroimaging feature selection with consistent
metric constraints for diagnosis of Alzheimer's disease. Medical Image
Analysis, 18(4), 425-435.
 Klöppel, S., et al. (2008). Automatic classification of MR scans in Alzheimer's
disease. Brain, 131(3), 681-689.
 Petersen, R. C., et al. (2014). Mild cognitive impairment: Ten years later.
Archives of Neurology, 69(3), 281-291.
 Shin, H. C., et al. (2016). Deep Convolutional Neural Networks for Computer-
Aided Detection: CNN Architectures, Dataset Characteristics and Transfer
Learning. IEEE Transactions on Medical Imaging, 35(5), 1285-1298.
 Esteva, A., et al. (2017). Dermatologist-level classification of skin cancer with
deep neural networks. Nature, 542(7639), 115-118.
 Rajpurkar, P., et al. (2017). CheXNet: Radiologist-Level Pneumonia Detection
on Chest X-Rays with Deep Learning. arXiv preprint arXiv:1711.05225.
 Gulshan, V., et al. (2016). Development and Validation of a Deep Learning
Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus
Photographs. JAMA, 316(22), 2402-2410.