Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1270

Alzheimers and Brain Tumor Detection Using

Deep Learning
Bandu Meshram1; Shubham Sudarshan More2; Ajinkya Padmakar Sagane3;
Datta Meghe4; Riddhesh Santosh Sarode5; Arpit Suryakant Lende6
1
Professor
1,2,3,4,5,6
Department of Computer Datta Meghe College of Engineering Airoli, India

Publication Date: 2025/05/01

Abstract: In several industries, such as manufacturing, construction, and the Accurate detection of brain tumors and
Alzheimer’s disease is essential for effective treatment and disease management. With the rapid progress in deep learning
technologies, the field of medical imaging—particularly the interpretation of brain scans—has seen remarkable
improvements. This research focuses on utilizing two well-established convolutional neural network (CNN) architectures,
VGG-19 and ResNet-50, for brain tumor classification, while employing a standard CNN model for detecting Alzheimer’s
disease.

VGG-19, characterized by its consistent and deep structure comprising 19 layers, is particularly effective in extracting
complex features due to its sequential convolutional layers. This makes it well-suited for identifying subtle patterns in MRI
images of the brain. In contrast, ResNet-50 incorporates residual connections within its 50-layer design, allowing the model
to mitigate issues like vanishing gradients and improving learning efficiency by enabling the network to focus on residual
mappings. This study compares both models to evaluate their accuracy, resilience, and computational efficiency in detecting
brain abnormalities.

Moreover, the research examines each model's ability to generalize across various datasets and tumor types, aiming to
provide insights into their clinical applicability. The results may contribute to refining current diagnostic techniques,
promoting earlier detection, and assisting in the development of advanced tools for accurate brain tumor diagnosis and
treatment planning. Integrating these models into healthcare systems could improve diagnostic accuracy and enhance
patient care outcomes. Alzheimer’s disease, the most prevalent form of dementia, leads to progressive memory loss, impaired
thinking, and behavioral changes. Its symptoms typically worsen over time, eventually hindering the ability to perform
everyday tasks. Dementia is a broad term describing a range of symptoms caused by cognitive decline, with Alzheimer’s
accounting for 60% to 80% of all cases. Vascular dementia, often following a stroke, is the second most common type, though
several other reversible conditions—like thyroid imbalances and vitamin deficiencies—can produce similar symptoms.In
this study, we use publicly available datasets for Alzheimer’s detection. The system employs deep learning models,
particularly CNN and ResNet, to analyze the data. The outcomes demonstrate the model's ability to accurately classify the
disease into categories such as mild, moderate, very moderate, and dementia, based on performance metrics like prediction
accuracy.

Keywords: Deep Learning, Convolutional Neural Networks (CNN), VGG-19, ResNet-50, Brain Tumor Detection, Alzheimer’s Disease, Medical
Image Analysis, MRI, Classification, Transfer Learning.

How to Cite: Prof Bandu Meshram ; Shubham Sudarshan More; Ajinkya Padmakar Sagane; Datta Meghe; Riddhesh Santosh
Sarode; Arpit Suryakant Lende (2025), Alzheimers and Brain Tumor Detection Using Deep Learning. International Journal of
Innovative Science and Research Technology, 10(4), 2073-2081. https://doi.org/10.38124/ijisrt/25apr1270

I. INTRODUCTION complexity of brain tumor detection arises from the diverse

types of tumors, their varied locations, and the subtlety of
Brain tumors represent a significant challenge in the early-stage symptoms. Traditionally, brain tumor detection
field of neurology and oncology, with their early detection relies on imaging techniques such as Magnetic Resonance
being crucial for effective treatment and improved patient Imaging (MRI), Computed Tomography (CT) scans, and
outcomes. The brain, being a critical organ, is highly sensitive positron emission tomography (PET) scans. However, these
to any pathological changes, and tumors can disrupt essential methods, while effective, often depend on the subjective
functions, leading to a range of neurological symptoms. The interpretation of radiologists, which can be influenced by

IJISRT25APR1270 www.ijisrt.com 2073

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1270
experience and training, leading to variability in diagnostic capability to learn high-level features from images. The depth
accuracy. and simplicity of VGG-19 allow it to capture intricate details
in brain MRI scans, making it a suitable candidate for tumor
ALZHEIMER’s disease (AD) is the leading cause of detection tasks.In contrast, ResNet-50, developed by
dementia and poses a significant social and economic Microsoft Research, introduces the concept of residual
challenge. It is responsible for more than half of all cases of learning through the use of residual blocks. These blocks
dementia. Over 50 million individuals currently suffer from include shortcut connections that bypass one or more layers,
dementia worldwide with a projected increase to 152 million mitigating the vanishing gradient problem and enabling the
by 2050. No cure for AD has been discovered, but there is network to learn residual mappings. ResNet-50’s architecture
intense effort to develop new clinical interventions that may allows it to be significantly deeper (50 layers) while
slow or halt the disease. Such interventions are aimed at early maintaining efficient training and performance. This
(including preclinical and prodromal) stages of the disease architecture has proven to be highly effective in handling
prior to extensive cell damage, when it is thought treatment is complex image classification tasks, making it a strong
more likely to be effective. Alzheimer’s disease and related contender for brain tumor detection.
dementias (ADRD) have become a major public health
concern in the United States. An estimated 5.6 million  Motivation
Americans aged 65 and older (10% of the US population) Early diagnosis of brain tumors and Alzheimer’s
were living with ADRD in 2019, and this number is expected disease remains a crucial yet challenging area in medical
to grow dramatically as the population continues to age. By practice due to the significant impact these conditions have
2025, the number of Americans aged 65 or older with ADRD on patient health. Timely identification is essential for
is expected to reach 7.1 million, nearly a 27% increase from initiating effective treatment and improving long-term
2019, and by 2050, this population is projected to be 13.8 outcomes. While traditional diagnostic techniques such as
million, with the highest growth among those in ADRD’s Magnetic Resonance Imaging (MRI) and Computed
advanced stage. Persons with ADRD require progressively Tomography (CT) scans are widely used to detect
extensive assistance in their daily lives, the majority of which neurological disorders, their effectiveness often hinges on the
is provided by family members, friends, and other unpaid expertise and subjective judgment of radiologists. This
caregivers. It is estimated that in 2018, American caregivers reliance can lead to variations in diagnosis, especially in early
of persons with ADRD provided 18.5 billion hours of stages where symptoms may be subtle or difficult to detect.
informal unpaid assistance, valued at $233.9 billion. Family
caregivers (hereafter “caregivers”) of persons with ADRD are This study is driven by the need to improve the accuracy
expected to make important care decisions for their family and reliability of diagnostic methods for brain tumors and
members with ADRD on a daily basis. However, these Alzheimer’s disease. It focuses on the use of deep learning—
caregivers report being unprepared for their roles and particularly convolutional neural networks (CNNs)—to
responsibilities, uninformed about care options, and enhance medical imaging analysis. CNNs have proven highly
unsupported by professionals in their decision making. effective in tasks involving image recognition and
classification, making them suitable candidates for use in the
 Role of Deep Learning in Medical Imaging healthcare domain.
In recent years, the field of medical imaging has been
revolutionized by the advent of deep learning technologies. By applying CNNs, the research aims to establish a
Deep learning, a subset of artificial intelligence (AI), utilizes more consistent and automated diagnostic approach. These
neural networks with multiple layers to learn and make models are capable of identifying intricate features and
predictions from complex data. Convolutional Neural patterns in brain scans that might not be easily noticed by the
Networks (CNNs) have emerged as a powerful tool in image human eye. This can help reduce the dependency on manual
analysis due to their ability to automatically learn spatial interpretation and lower the chances of misdiagnosis or
hierarchies of features from images. CNNs have variability between radiologists’ assessments.
demonstrated remarkable performance in various medical
imaging tasks, including tumor detection, segmentation, and Furthermore, incorporating CNNs into diagnostic
classification. These networks can enhance diagnostic workflows could enable earlier detection, even before
accuracy by identifying patterns and anomalies that might be noticeable symptoms arise. This early intervention can play a
missed by the human eye, thus providing a valuable pivotal role in treatment planning and patient care. An
augmentation to traditional diagnostic methods. automated system based on CNNs could also act as a
supportive tool for radiologists, offering a reliable second
VGG-19 and ResNet-50: Architectures and opinion and enhancing confidence in clinical decisions.
Applications Among the various CNN architectures
developed, VGG-19 and ResNet-50 are prominent due to II. LITERATURE SURVEY
their innovative designs and effective performance in image
classification tasks. VGG-19, introduced by the Visual [1]The article "Deep Learning Approaches for Brain
Geometry Group (VGG) at the University of Oxford, is a Tumor Detection in MRI Images: A Comprehensive Survey"
deep CNN characterized by its uniform architecture of 19 provides a thorough examination of various deep learning
layers. Its design is notable for using small convolutional models applied to MRI-based brain tumor detection. It
filters (3x3) and pooling layers, which contribute to its explores key architectures like VGGNet, ResNet,

IJISRT25APR1270 www.ijisrt.com 2074

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1270
EfficientNet, and AlexNet, detailing their convolutional imaging, can significantly improve diagnostic accuracy.
layers, residual blocks, and attention mechanisms. The paper MobileNet-BT's fast convergence and superior accuracy
emphasizes image preprocessing methods such as make it ideal for enhancing brain tumor diagnosis, offering a
augmentation and normalization, as well as training substantial improvement in clinical applications.[6] Deep
optimizations like transfer learning. Performance metrics like Learning Approaches for Brain Tumor Detection in MRI
accuracy, F1-score, and computational complexity are Images: A Comprehensive Survey. The article "Deep
compared to highlight model efficiency in clinical tumor Learning Approaches for Brain Tumor Detection in MRI
diagnosis.[2] The report explores non-invasive brain tumor Images: A Comprehensive Survey" provides a thorough
detection using MRI images, employing deep learning (DL) examination of various deep learning models applied to MRI-
and machine learning (ML) techniques. It uses a dataset of based brain tumor detection. It explores key architectures like
3264 MRI images to classify three tumor types—glioma, VGGNet, ResNet, EfficientNet, and AlexNet, detailing their
meningioma, and pituitary tumors—along with healthy convolutional layers, residual blocks, and attention
brains. Two models were developed: a 2D Convolutional mechanisms. The paper emphasizes image preprocessing
Neural Network (CNN) and a convolutional autoencoder. methods such as augmentation and normalization, as well as
The 2D CNN achieved a high training accuracy of 96.47%, training optimizations like transfer learning. Performance
outperforming the autoencoder (95.63%) and traditional ML metrics like accuracy, F1-score, and computational
methods like K-Nearest Neighbors (KNN), which had the complexity are compared to highlight model efficiency in
highest accuracy among ML models at 86%. The study clinical tumor diagnosis.[7] This paper discusses the
demonstrates the effectiveness of the 2D CNN, highlighting application of deep learning architectures in healthcare
its potential use in clinical settings for accurate brain tumor diagnostics, focusing on disease detection from medical
detection.[3] The report "MRI-based brain tumor detection images such as X-rays, MRIs, and CT scans. The authors
using convolutional deep learning methods and chosen evaluate various models, including CNNs (Convolutional
machine learning techniques" focuses on the use of deep Neural Networks) and their derivatives like VGG-19 and
learning (DL) and machine learning (ML) for non-invasive ResNet-50, to determine their accuracy in diagnosing
brain tumor detection using MRI images. The study utilizes a diseases such as cancer and cardiovascular issues. The
dataset of 3264 MRI images to classify glioma, meningioma, research highlights improvements in prediction accuracy and
pituitary gland tumors, and healthy brains. Two DL models discusses the potential of AI to revolutionize early-stage
were developed: a 2D Convolutional Neural Network (CNN) disease detection.[8] The paper investigates AI applications
and a convolutional autoencoder. The 2D CNN, comprising in cybersecurity, emphasizing the importance of machine
eight convolutional and four pooling layers, achieved a learning and deep learning techniques to enhance the
training accuracy of 96.47% and a validation accuracy of detection of cyber threats. By using anomaly detection
93.44%. The autoencoder also performed well with a training methods and network traffic analysis, the authors present
accuracy of 95.63% and validation accuracy of 90.93%. Both models capable of identifying patterns associated with
models demonstrated high recall rates and ROC curves close malware, phishing, and DDoS attacks. They also explore the
to 1, making them reliable for brain tumor detection.[4] challenges in creating systems that can adapt to evolving
Robust brain tumor classification by fusion of deep learning cyber threats and discuss how AI can help automate threat
and channel-wise attention mode approach. The report detection with higher precision and efficiency than traditional
focuses on improving brain tumor classification using deep systems.[9] This research focuses on the integration of AI
learning, specifically the ResNet101 model combined with a into supply chain management, particularly in predictive
Channel-wise Attention Mechanism (CWAM). This modeling to forecast demand, optimize inventory, and
approach aims to enhance the accuracy and efficiency of mitigate risks such as delays and stock shortages. The authors
diagnosing brain tumors from MRI scans, traditionally reliant analyze various AI-based algorithms like Random Forest,
on radiologists. The ResNet101-CWAM model was trained Decision Trees, and neural networks to improve the accuracy
using preprocessed MRI data and achieved an impressive of supply chain forecasting and decision-making processes.
accuracy of 99.83%, outperforming existing methods. The They highlight how AI can help companies enhance their
attention mechanism allowed the model to focus on key operational efficiency and adapt to changes in global supply
features, improving classification. The study suggests chains, especially during disruptions such as those caused by
integrating this model into clinical software for better the COVID-19 pandemic.[10] This article presents a novel
decision-making in treatment planning. While the results are solution for detecting Alzheimer’s disease and brain tumors
promising, the authors highlight the need for optimizing the using machine learning (ML) and Convolutional Neural
model’s computational complexity and recommend further Networks (CNNs). It highlights the critical need for early
research on diverse datasets to improve the system further.[5] diagnosis due to the increasing prevalence of these
Deep Learning in Medical Image Classification from MRI- neurological disorders. The proposed approach integrates
based Brain Tumor Images. The paper investigates brain multi-modal data fusion (e.g., combining MRI, PET, and CT
tumor classification using MRI images with four pre-trained scans with clinical information such as patient demographics
models. MobileNetV2, ResNet-18, EfficientNet-B0, and and genetic markers) to enhance diagnostic accuracy. Key
VGG16—and introduces MobileNet-BT. This new model, techniques include the use of advanced CNN architectures
built on MobileNetV2, achieved a remarkable 99.24% like ResNet and DenseNet to improve feature extraction from
accuracy and F1-score, surpassing the other models in medical images, attention mechanisms for prioritizing key
efficiency and performance. The research emphasizes how features, and transfer learning to overcome limited medical
fine-tuning pre-trained models, specifically for medical datasets. The article emphasizes the potential of combining

IJISRT25APR1270 www.ijisrt.com 2075

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1270
CNNs with Recurrent Neural Networks (RNNs) for temporal sensors, can be used to optimize crop production, monitor soil
modeling, allowing for the prediction of disease progression. health, and reduce water usage. They highlight the potential
The framework aims for clinical applicability by improving of AI to enhance decision-making in agriculture, enabling
the robustness and accuracy of automated diagnosis.[11] This farmers to use data-driven insights for improving yield,
paper provides a comprehensive overview of brain tumor managing pests, and minimizing environmental impact. The
detection and classification using computational intelligence paper also touches on the challenges of AI adoption in rural
techniques. It reviews existing methods, focusing on areas due to infrastructure limitations.
challenges like tumor variability in position, structure, and
size. The study covers a wide range of techniques, including III. PROBLEM STATEMENT
traditional image processing and machine learning methods,
such as Leksell Gamma Knife and Radioactive beams, as well The project seeks to tackle the challenges of early and
as more recent advances in deep learning (DL) and transfer accurate detection of brain tumors and Alzheimer's disease by
learning (TL). Notable DL models discussed include 3D leveraging deep learning models, including VGG-19,
CNNs, DenseNet, and Mask-RNNs, which are applied for the ResNet-50, and 2D CNN architectures, to automate MRI scan
segmentation and classification of brain tumors. The paper analysis, improve diagnostic precision, and overcome
highlights that MRI is the primary imaging modality used for limitations such as manual interpretation errors and vanishing
brain tumor detection and discusses the performance of gradients in neural networks. By utilizing transfer learning
various models on standard datasets, like BRATS. The use of techniques, the pre-trained models can effectively adapt to
pre-processing, feature extraction, and machine learning medical imaging tasks, reducing the need for large amounts
classifiers is covered in detail, with suggestions for future of labeled data and enabling faster convergence. The
research on improving accuracy through hybrid models and integration of advanced image preprocessing techniques,
advanced pre-processing techniques.[12] This article such as gray-level co-occurrence matrix (GLCM) for texture
explores brain tumor and Alzheimer’s detection using deep analysis and feature extraction, ensures that the models can
learning models, focusing on the classification of brain capture both global and local features essential for
tumors and the early detection of Alzheimer’s disease. The distinguishing between normal and pathological brain
authors developed an Alzheimer’s detection model that conditions.
classifies patients based on three levels of severity (very mild,
mild, moderate dementia), and a brain tumor detection model IV. METHODOLOGY
that classifies three types of tumors: meningioma, pituitary,
and glioma. Using MRI scans and CNN architectures, the  Dataset Collection
Alzheimer’s model achieved an accuracy of 98.37%, while MRI images of brain tumors and Alzheimer’s disease
the brain tumor detection model reached 97.16% accuracy. were sourced from publicly available repositories such as
The article also emphasizes segmentation techniques for Kaggle and other medical imaging databases. The datasets
localizing tumors in brain MRIs. The study concludes that were selected to ensure diversity in terms of tumor types,
deep learning models are highly effective for both detection stages, and patient demographics, providing a reliable
and classification tasks, and highlights the potential for these foundation for model training and evaluation.
models to be further optimized for real-time medical
diagnosis applications.[13] This paper focuses on the use of  Image Preprocessing
AI and machine learning for improving natural language Preprocessing steps included resizing the images to a
processing (NLP) tasks. The authors explore advanced NLP uniform resolution (e.g., 300x300 pixels) and converting
techniques such as BERT (Bidirectional Encoder them to grayscale to simplify the data and reduce
Representations from Transformers) and GPT (Generative computational load. Image normalization was also applied to
Pre-trained Transformer) for applications in text generation, standardize the pixel intensity distribution, thereby enhancing
sentiment analysis, and language translation. They examine feature detection during model training.
the effectiveness of these models in handling large datasets,
improving human-computer interaction, and developing  Image Resizing:
more accurate, context-aware systems for real-world All input MRI and CT scan images were resized to a
applications like chatbots and virtual assistants.[14] This fixed resolution of 300x300 pixels. This standardization
research explores the integration of AI in healthcare, ensures that the convolutional neural networks (VGG-19,
particularly in predicting disease outbreaks using AI-driven ResNet-50, and CNN) can process the data in batches without
predictive analytics. The authors investigate how data from dimensional inconsistencies.
healthcare systems, combined with environmental and social
factors, can be leveraged to build models that predict the  Grayscale Conversion:
spread of infectious diseases like COVID-19 and influenza. Since color information is not essential for identifying
The paper also discusses the role of AI in resource allocation brain abnormalities, RGB images were converted to
during outbreaks, highlighting the importance of real-time grayscale. This reduces the number of input channels from
decision-making in managing healthcare resources and three to one, significantly decreasing memory usage and
reducing the spread of diseases.[15] The paper analyzes the processing time, while preserving critical structural details
application of AI in the field of agriculture, specifically in necessary for diagnosis.
precision farming. The authors discuss how AI technologies,
such as machine learning models and IoT (Internet of Things)

IJISRT25APR1270 www.ijisrt.com 2076

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1270
VGG-19, known for its deep and uniform architecture
with 19 weight layers, was employed due to its effectiveness
in capturing intricate spatial hierarchies within MRI images.
The use of small 3x3 convolution filters in successive layers
allowed the model to focus on fine-grained features, making
it particularly effective for detecting subtle anomalies in brain
tissue.

ResNet-50, a 50-layer deep residual network,

introduced the concept of skip connections or residual blocks,
which help the network learn identity mappings. This
approach effectively mitigates the vanishing gradient
problem during backpropagation, thus enabling deeper
networks to train efficiently without degradation in
performance. ResNet-50 is especially useful for handling
complex datasets with a wide variety of tumor and tissue
textures.

Fig 1 Workflow Diagram A custom CNN model was also designed and optimized
specifically for Alzheimer’s detection. This model included
 Feature Extraction multiple convolutional and pooling layers followed by dense
Both statistical and textural features were extracted layers, using ReLU activations and softmax for final
from the preprocessed images. Techniques such as mean, classification. While less deep than the VGG-19 and ResNet-
median, variance, and Gray-Level Co-occurrence Matrix 50 models, it offered flexibility and computational efficiency
(GLCM) were employed to derive meaningful insights. for the Alzheimer's-specific dataset.
GLCM features like contrast, correlation, energy, entropy,
and homogeneity helped capture spatial texture relationships, All models were trained using transfer learning, where
essential for distinguishing healthy and abnormal brain pretrained weights from large datasets like ImageNet were
regions. adapted to the medical imaging task. Fine-tuning of upper
In addition to GLCM, histogram-based features were layers was performed to tailor the models to domain-specific
analyzed to understand pixel intensity distribution, which features, ensuring improved accuracy even with limited
assists in identifying abnormalities in grayscale values. Edge labeled data.
detection filters were also utilized to highlight boundaries of
tumor-affected areas. These extracted features were then The dataset was divided into training (70%) and testing
flattened into a numerical vector format suitable for feeding (30%) subsets using a stratified approach to preserve class
into the neural network models. The combination of both distribution. The models were compiled using categorical
global statistical and local texture-based features ensured cross-entropy as the loss function and optimized using the
robust input representation for accurate classification and Adam optimizer with learning rate scheduling and dropout
prediction. for regularization. Training was conducted over multiple
epochs with batch normalization to stabilize and accelerate
 Model Development convergence.
To achieve accurate classification of brain tumors and
Alzheimer’s disease, this study implemented three deep Performance metrics such as accuracy, precision, recall,
learning models—VGG-19, ResNet-50, and a custom- and F1-score were tracked during training, and early stopping
designed Convolutional Neural Network (CNN). Each model was used to prevent overfitting. This multi-model strategy
was carefully chosen or developed to exploit its strengths in allowed for comparative analysis and identification of the
medical image analysis. most suitable architecture for medical diagnosis.

IJISRT25APR1270 www.ijisrt.com 2077

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1270

Fig 2 VGG -19

VGG-19 is a deep convolutional neural network known for its ability to handle large-scale image
consisting of 19 weight layers, primarily using small 3x3 classification tasks. VGG-19 is effective in recognizing
convolution filters and 2x2 max-pooling layers. Its subtle variations in MRI scans due to its depth and fine-
architecture follows a very uniform and simple structure, grained feature extraction. Despite its high accuracy, it
making it easier to implement and tune. The model extracts requires significant computational resources due to the large
hierarchical features from images, starting from edges in the number of parameters.
initial layers to complex patterns in deeper layers. It is widely

Fig 3 Convolutional Neural Network (CNN)

CNN is a class of deep learning models designed classification. CNNs automatically learn to extract relevant
specifically for analyzing visual data. It processes input spatial information from raw images without manual feature
images through multiple layers of convolution, activation engineering. They are widely used in medical image analysis
(usually ReLU), and pooling, which help in detecting features due to their high accuracy and adaptability. By stacking
such as edges, textures, and shapes. These features are then layers, CNNs can learn increasingly complex patterns to
passed through fully connected layers to perform distinguish between healthy and diseased tissues.

IJISRT25APR1270 www.ijisrt.com 2078

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1270

Fig 4 ResNet50 Model Architecture

ResNet-50 is a 50-layer deep neural network that During the classification process, the input image
incorporates residual blocks, allowing it to train much deeper passed through the network’s layers where important features
architectures without performance degradation. The key were automatically extracted. Convolution and pooling
innovation in ResNet is the use of shortcut or skip operations helped in reducing the image's dimensionality
connections that pass the input of a layer directly to a deeper while preserving crucial information needed for accurate
layer, helping to maintain gradient flow during recognition. After feature extraction, the information was
backpropagation. This architecture addresses the vanishing aggregated through fully connected layers, preparing the data
gradient problem commonly seen in very deep networks. for final classification. The softmax layer then generated the
ResNet-50 is highly effective in complex image recognition probabilities for each possible class, and the model selected
tasks, offering improved accuracy and faster convergence. In the class with the maximum probability as the final output.
medical imaging, it is particularly useful for identifying
detailed patterns in high-resolution scans and handling In the case of Alzheimer’s detection, the system not
diverse datasets. only identified the presence of the disease but also assessed
the severity of its progression. By analyzing structural brain
V. CLASSIFICATION AND PREDICTION changes, such as hippocampal shrinkage and cortical atrophy,
the models could classify patients into different stages of the
After the deep learning models (VGG-19, ResNet-50, disease. This detailed classification was critical in helping
and a custom CNN) were effectively trained using the healthcare professionals provide early diagnosis and plan
preprocessed and labeled medical images, the next important effective treatment strategies tailored to each patient’s
phase involved classifying new and unseen MRI scans. The condition.
main objective was to determine whether a brain MRI
indicated the presence of a tumor or Alzheimer’s disease and, After obtaining predictions, the model outputs were
if applicable, accurately classify the type or stage of the compared with the ground truth labels from the test dataset.
condition. To evaluate the performance and reliability of the
For brain tumor detection, the models were trained to classification results, metrics such as accuracy, precision,
distinguish between normal brain scans and those exhibiting recall, and F1-score were calculated. Accuracy measured the
tumor-affected areas. In more complex cases, the models overall percentage of correct predictions, while precision
were capable of further classifying different tumor types, assessed how many of the positive predictions were actually
such as glioma, meningioma, and pituitary tumors, depending correct. Recall determined the model’s ability to detect all
on the dataset that was used. Similarly, for Alzheimer’s actual positive cases, and the F1-score provided a balanced
disease detection, the models categorized brain images into evaluation by combining both precision and recall. Together,
various stages, including mild cognitive impairment, these performance indicators offered a comprehensive
moderate Alzheimer’s, severe Alzheimer’s, and full understanding of how well the models could classify brain
dementia, allowing for more precise clinical insights. tumors and Alzheimer's disease, highlighting their potential
to support clinical decision-making and early intervention.
To achieve this, each model’s final output layer was
constructed to perform multi-class classification through the
use of the softmax activation function. Softmax transforms
the model’s raw output scores into a probability distribution
across all possible categories. This approach allowed the
model to predict the most likely class by selecting the label
with the highest probability, offering a clear and confident
final decision for each input image.

IJISRT25APR1270 www.ijisrt.com 2079

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1270
Performance evaluation metrics such as accuracy,
precision, recall, and F1-score indicated that the proposed
models maintained a strong balance between correctly
identifying positive cases and minimizing misclassifications.
These results support the notion that deep learning can
significantly enhance early detection and diagnosis of
neurological disorders, leading to better treatment planning
and improved patient outcomes.

Moreover, the successful classification of Alzheimer’s

stages based on subtle brain structure changes highlights the
potential for deep learning to assist in preclinical diagnosis,
which is critical for managing progressive conditions.
Integrating such automated systems into clinical workflows
could serve as a valuable decision-support tool for healthcare
professionals, reducing diagnostic variability and aiding in
early interventions.

Overall, the research validates the potential of deep

learning in revolutionizing medical imaging analysis. Future
work could focus on expanding the models to handle larger,
more diverse datasets, applying real-time processing
techniques, and refining the architectures to further enhance
clinical applicability and scalability across different
healthcare environments.

REFERENCES

[1] Saikat Islam Khan, Anichur Rahman,a Tanoy

Debnath, Md. Razaul Karim, Mostofa Kamal Nasir,a
Shahab S. Band, Amir Mosavi “Accurate brain tumor
detection using deep convolutional neural
networkMd”, 2022 Aug 27, Computational and
Structural Biotechnology Journal, Vol. 20.
Fig 5 Flowchart of Model Generation [2] Soheila Saeedi, Sorayya Rezayi, Hamidreza
Keshavarz & Sharareh R. Niakan Kalhori “MRI-based
VI. CONCLUSION brain tumor detection using convolutional deep
learning methods and chosen machine learning
In this study, deep learning models such as VGG-19, techniques” 23 January 2023, Scientific Reports, Vol.
ResNet-50, and a custom-designed Convolutional Neural 13.
Network were successfully implemented for the classification [3] Vinodhini Mani, 1 S. R. Srividhya, 1 Osamah Ibrahim
and prediction of brain tumors and Alzheimer’s disease using Khalaf, 2 and Carlos Andrés Tavera Romero “Early-
MRI images. The research demonstrated that these models Stage Alzheimer's Disease Prediction Using Machine
could effectively distinguish between healthy brain scans and Learning ModelsC. Kavitha, corresponding” 2022
those exhibiting pathological changes, and could further Mar 3, Healthcare, Vol. 10.
classify tumors and Alzheimer’s stages with a high degree of [4] Balamurugan A.G, Saravanan Srinivasan, Mohd Asif
accuracy. The application of the softmax activation function Shah “Robust brain tumor classification by fusion of
in the final layer enabled multi-class classification, allowing deep learning and channel-wise attention mode
the models to deliver reliable probability-based predictions. approac0h” 17 June 2024, IEEE Access, Vol. 12
[5] Xiaoyi Liu, Zhuoyue Wang “Deep Learning in
The findings emphasized the effectiveness of deep Medical Image Classification from MRI- based Brain
learning architectures in extracting intricate features from Tumor Images” 1 Aug 2024, IEEE Transactions on
complex medical images, outperforming traditional Medical Imaging, Vol. 43.
diagnostic methods that rely heavily on manual [6] Rahul Namdeo Jadhav, 2 Dr.G.Sudhaghar. “Deep
interpretation. Particularly, the VGG-19 model exhibited Learning Approaches for Brain Tumor Detection in
strength in capturing detailed spatial hierarchies, while the MRI Images: A Comprehensive Survey”, 16/01/2024,
ResNet-50 model proved robust in training efficiency due to Journal of King Saud University – Computer and
its residual learning technique. The custom CNN model Information Sciences, Vol. 36.
tailored for Alzheimer’s detection also provided efficient and [7] Vijeeta Patil, Manohar Madgi and Ajmeera Kiran,”
accurate classification despite being less computationally Early prediction of Alzheimer’s disease using
intensive. convolutional neural network: a review” , 2022,

IJISRT25APR1270 www.ijisrt.com 2080

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1270
Journal of Neural Engineering, Vol. 19.
[8] Omer Asghar Dara 1, Jose Manuel Lopez-Guede 1,
Hasan Issa Raheem 1, Javad Rahebi 2, Ekaitz Zulueta
1 and Unai Fernandez-Gamiz,Alzheimer’s, “Disease
Diagnosis Using Machine Learning:A Survey,” 18
July 2023, Sensors, Vol. 23.
[9] Rajeshwari dharavath, Kattula Shyamala, “MRI Image
Based Relatable Pixel Extraction with Image
Segmentation for Brain Tumor Cell Detection Using
Deep Learning Model’, 2021, International Journal of
Imaging Systems and Technology, Vol. 31.32
[10] Shaik firoz1 N.M.V. Nirmala, “Detection and
classification of brain tumors With Alzheimer using
deep learning and Convolutional neural networks”,
March 2021, Journal of Intelligent Systems, Vol. 30.
[11] Rahmeh Ibrahim 1, Rawan Ghnemat 1 and Qasem Abu
Al-Haija 2, “Improving Alzheimer’s Disease and
Brain Tumor Detection Using Deep Learning with
Particle Swarm Optimization”, 28 July 2023,
Electronics, Vol. 12.
[12] Vasco Sá Diogo1,2 Hugo Alexandre Ferreira1†, Diana
Prata “Early diagnosis of Alzheimer’s disease using
machine learning: a multi-diagnostic, generalizable
approach”, 2023, Frontiers in Aging Neuroscience,
Vol. 15.
[13] Poshitha M, Nazia Sultana A, Hemathanmaya S S,
Mahendra A, J C Vasantha Kumar, Brain “Tumor and
Alzheimer’s Detection using Deep Learning” June
2023, IEEE Journal of Biomedical and Health
Informatics, Vol. 27.
[14] Nag Shivani Puram, Detecting Alzheimer’s “Disease
and brain tumors using machine learning and
convolutional neural networks (CNNs)”, 18
September 2024, Computational Intelligence, Vol. 40.
[15] Shubhangi Solanki , Uday Pratap Singh , (Member,
Ieee), Siddharth Singh Chouhan , And Sanjeev Jain
, “Brain Tumor Detection and Classification Using
Intelligence Techniques”, An Overview 6 Feb 2023,
IEEE Access, Vol. 11.

IJISRT25APR1270 www.ijisrt.com 2081