A Deep Learning Model for Effective Brain Tumor

Prediction using MRI Images

A project report submitted in partial fulfillment

of the requirements for the degree of

Bachelor of Technology
in

Electronics & Computer Engineering

by

SUSHANTH B S (21BLC1260)

HARSHAVISHAL S (21BLC1349)

HARSAVARDHANAN K V (21BLC1556)

Under the Guidance of

DR. SUKRITI

School of Electronics Engineering,

Vellore Institute of Technology, Chennai,

Vandalur-Kelambakkam Road,

Chennai - 600127, India.

April 2025
 Declaration

I hereby declare that the report titled A Deep Learning Model for effective Brain
Tumor Prediction using MRI Images submitted by us to the School of Electron-
ics Engineering, Vellore Institute of Technology, Chennai in partial fulfillment of the
requirements for the award of Bachelor of Technology in Electronics and Com-
puter Engineering is a bona-fide record of the work carried out by me under the
supervision of Dr. Sukriti .

I further declare that the work reported in this report, has not been submitted and will
not be submitted, either in part or in full, for the award of any other degree or diploma
of this institute or of any other institute or University.

Sign:

Name & Reg. No.:

Date:

i
 School of Electronics Engineering

Certificate
This is to certify that the project report titled A Deep Learning Model for ef-
fective Brain Tumor Prediction using MRI Images submitted by Sushanth
B S (21BLC1260) ,Harshavishal S (21BLC1349) & Harsavardhanan K V
(21BLC1556) to Vellore Institute of Technology Chennai, in partial fulfillment of the
requirement for the award of the degree of Bachelor of Technology in Electronics
and Computer Engineering is a bona-fide work carried out under my supervision.
The project report fulfills the requirements as per the regulations of this University and
in my opinion meets the necessary standards for submission. The contents of this report
have not been submitted and will not be submitted either in part or in full, for the award
of any other degree or diploma and the same is certified.

Supervisor Head of the Department

Signature: .................... Signature: ....................

Name: .................... Name: ....................

Date: Date:

Examiner 1 Examiner 2

Signature: . . . . . . . . . . . . . . . . . . . . . . . . Signature: . . . . . . . . . . . . . . . . . . . . . . . .

Name: . . . . . . . . . . . . . . . . . . . . . . . . . . . . Name: . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Date: Date:

(Seal of the School)

 Abstract

Brain tumors are significant risk to life, making it vital to quickly and accurately di-
agnose them for successful treatment can increase the suvival rate of a human. So a
method based on deep learning which aims in detecting brain tumors through MRI
images, introducing a Dual Attention Residual Network(DARESNET) module that in-
corporates both spatial and channel attention mechanisms along with residual connec-
tions. This enhances the models ability to refine most informative features in brain
MRI scans leading to more accurate and robust tumor classification. Moreover DARES-
NET is integrated into several state of the art CNN backbones like MobileNetV3Large,
ConvNextBase, ResNet50, DenseNet201 and VGG16 demonstrating its flexibility and
adaptability across different deep learning models.

The models are being trained and tested on three publicly available brain tumor datasets,
ensuring robustness and generalizability, achieving highest accuracy of 99.73% on Dataset-
1, 99.03% on Dataset-2 and 99.26% on Dataset-3. This model enhances tumor identifi-
cation and improves accuracy relative to standard CNN architectures.The experimental
findings confirm the models capability in effectively differentiating MRI scans that fea-
ture tumors from those that do not. These results tells the role of deep learning in
enabling automated and dependable identification of brain tumors.
 Acknowledgements
We wish to express our sincere thanks and deep sense of gratitude to our project guide,
Dr.Sukriti, Assitant Professor, School of Electronics Engineering, for her consistent en-
couragement and valuable guidance offered to us in a pleasant manner throughout the
course of the project work.

We are extremely grateful to Dr. Ravishankar A, Dean Dr. Reena Monica, Associate
Dean (Academics) & Dr. John Sahaya Rani Alex, Associate Dean (Research) of the
School of Electronics Engineering, VIT Chennai, for extending the facilities of the School
towards our project and for their unstinting support. We express our thanks to our Head
of the Department Dr. Annis Fathima A for her support throughout the course of this
project.

We also take this opportunity to thank all the faculty of the School for their support and
their wisdom imparted to us throughout the course. We thank our parents, family, and
friends for bearing with us throughout the course of our project and for the opportunity
they provided us in undergoing this course in such a prestigious institution.

iv
Contents

Declaration i

Certificate ii

Abstract iii

Acknowledgements iv

List of Figures vii

List of Tables viii

1 Introduction 1

2 Literature Survey 4
2.1 Review of Binary Class Brain Tumor Classification . . . . . . . . . . . . . 4
2.2 Review of Multi-Class Brain Tumor Classification . . . . . . . . . . . . . . 5
2.3 Research Gaps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3 Methodology 11
3.1 Database Used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.1.1 Dataset-1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.1.2 Dataset-2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1.3 Dataset-3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 Image Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2.1 Creating Dataframe . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2.2 Convert DataFrame into TensorFlow Dataset . . . . . . . . . . . . 14
3.2.3 Image Resizing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2.4 Image Normalization . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3 Image Augmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3.1 Image Flipping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3.2 Image Brightness Adjustment . . . . . . . . . . . . . . . . . . . . . 16
3.3.3 Image Saturation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.3.4 Image Contrast . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

v
Contents vi

3.4 Image Batch & Prefetching . . . . . . . . . . . . . . . . . . . . . . . . . . 18

3.5 Pretrained CNN Backbone Models . . . . . . . . . . . . . . . . . . . . . . 18
3.6 Dual Attention Residual Network . . . . . . . . . . . . . . . . . . . . . . . 19
3.6.1 Channel Attention . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.6.2 Spatial Attention . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.6.3 Residual Connections . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.7 GlobalAveragePooling2D . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.8 Dense (128, ReLU Activation) . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.9 Dense (n, Softmax Activation) . . . . . . . . . . . . . . . . . . . . . . . . 24
3.10 Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

4 Results and Discussions 25

4.1 Dataset Splitting & Hyperparameters Tuning . . . . . . . . . . . . . . . . 25
4.2 Model Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2.1 Accuracy and Loss Plots . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2.2 Confusion Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2.3 Precision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2.4 Recall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2.5 F1-Score . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2.6 Confidence Scores . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.3 Comparison of CNN-DARESNET VS CNN Architectures using Dataset 1 28
4.4 Classification Results of CNN-DARESNET using Dataset 2 . . . . . . . . 32
4.5 Classification Result of CNN-DARESNET using Dataset 3 . . . . . . . . . 35
4.6 Comparison on Existing and Previous Study . . . . . . . . . . . . . . . . . 38
4.6.1 Comparison with Binary Classification Models . . . . . . . . . . . 39
4.6.2 Comparison with Multi-Class Classification Models . . . . . . . . . 39

5 Conclusion and Future Scope 44

6 Appendix 45

7 Bibliography 47

8 Biodata 50
List of Figures

3.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2 Dataframe Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3 Tensorflow Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.4 Image Resizing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.5 Image Normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.6 Image Flipping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.7 Image Brightness Adjustment . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.8 Image Saturation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.9 Image Contrast . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.10 CNN Backbone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.11 Channel Attention Module . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.12 Spatial Attention Module . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.13 Residual Connections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

4.1 Training Accuracy Plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

4.2 Training Accuracy Plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.3 Confidence Scores of the model which achieved highest accuracy . . . . . 29
4.4 Training vs Validation Loss Plot . . . . . . . . . . . . . . . . . . . . . . . 30
4.5 Training vs Validation Accuracy Plot . . . . . . . . . . . . . . . . . . . . . 31
4.6 Confusion Matrix of the model . . . . . . . . . . . . . . . . . . . . . . . . 32
4.7 Training Accuracy Plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.8 Confidence Scores of the model which achieved highest accuracy . . . . . 33
4.9 Training vs Validation Loss Plot . . . . . . . . . . . . . . . . . . . . . . . 34
4.10 Training vs Validation Accuracy Plot . . . . . . . . . . . . . . . . . . . . . 34
4.11 Confusion Matrix of the model . . . . . . . . . . . . . . . . . . . . . . . . 35
4.12 Training Accuracy Plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.13 Confidence Scores of the model which achieved highest accuracy . . . . . 36
4.14 Training vs Validation Loss Plot . . . . . . . . . . . . . . . . . . . . . . . 37
4.15 Training vs Validation Accuracy Plot . . . . . . . . . . . . . . . . . . . . . 37
4.16 Confusion Matrix of the model . . . . . . . . . . . . . . . . . . . . . . . . 38

vii
List of Tables

3.1 Dataset Distribution of Dataset-1 . . . . . . . . . . . . . . . . . . . . . . . 12

3.2 Dataset Distribution of Dataset-2 . . . . . . . . . . . . . . . . . . . . . . . 13
3.3 Dataset Distribution of Dataset-3 . . . . . . . . . . . . . . . . . . . . . . . 13

4.1 Optimized Hyperparameters . . . . . . . . . . . . . . . . . . . . . . . . . . 26

4.2 Classification Report and (Testing & Validation ) Accuracy . . . . . . . . 29
4.3 Classification Report and (Testing Validation ) Accuracy . . . . . . . . . 33
4.4 Classification Report and (Testing Validation ) Accuracy . . . . . . . . . 35
4.5 Comparison of Brain Tumor Detection Methods . . . . . . . . . . . . . . . 41
4.6 Comparison of review on binary class brain tumor classification using
Dataset-2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.7 Comparison of review on multi-class brain tumor classification using Dataset-
3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

viii
Chapter 1

Introduction

A brain tumor can be defined as the uncontrolled and abnormal development of cells
inside the brain. This extra growth inside the brain might cause more weight inside the
skull causing serious complications and harming the brain. The meaning and suggestions
of brain tumors can be different depending on whether it is benign or malignant. Benign
ones are non cancerous which is not being spread to other parts of the brain or body.
Malignant ones are cancerous which develops exponentially have chances to spread to
other parts of the body.

There are two fundamental brain tumor types primary and secondary. Primary brain
tumors starts form the brain. They may be either cancerous or non cancerous with
faster developing cancerous cells. Secondary brain tumors starts from other parts of the
body such as the breasts, lungs, kidneys or skin but afterwards it spreads to the brain
through blood. These types are always cancerous and never generous.

Primary tumors can be further categorized to glial and non glial tumors. Glial tumors or
gliomas arise inside the glial cells. This cells supports the nervous system by surrounding
and holding neurons in place providing nutrients and oxygen to the nerve cells, removing
dead neurons and insulating neurons from each other. Cases of gliomas are Astrocytoma,
Oligodendroglia tumors and Glioblastomas. Non glial tumors develops in structures of
the brain such as blood vessels, nerves, meninges or glands. Cases includes meningiomas,
pituitary tumors, schwannomas, pineal gland tumors and craniopharyngiomas. Among
more than 120 brain tumor types glioma, meningioma and pituitary have been detected
and identified as the foremost common ones.

Brain tumor detection is very complicated and difficult due to the size, shape, location
and type of tumor in the brain. However, if the tumor is diagnosed and treated early
in the tumor growth process, the chance of patients treatment is very high. Therefore,

1
Chapter 1. Introduction 2

the treatment of tumor depends on the early and timely diagnosis of the tumor. The
diagnosis is usually done by a medical examination using different experimental proce-
dures, including biopsy, Cerebrospinal fluid analysis, Skull X-ray analysis and imaging
tests . Nowadays image modalities are becoming more popular among radiologists since
they are more accurate and introduce much less risk to patients. Imaging techniques
like Magnetic Resonance Imaging (MRI), Computerized Tomography (CT) and Positron
Emission Tomography (PET) scans are used to treat brain tumors. Among them MRI
images play a crucial role in the early treatment of the tumor and its classification. Be-
cause of its high pixel values in soft tissues and the non-invasiveness attribute without
causing much radiation contamination, MRI images are given more priority compare
to other medical imaging techniques. In MRI there are different types of imaging scan
technologies that are available to generate various types of tissue images that offer struc-
tural evidence for the diagnosis to cure of brain tumors. Due to its several advantages,
availability and imaging techniques with the high resolution image quality of even soft
tissue, we use MRI for the detection of brain tumors.

Physicians primarily rely on conventional inspection when identifying brain tumors from
MRI images. However, manually interpreting these images has become time-consuming,
chaotic, and usually wrong as the number of patients has increased. To overcome this
limitation, a computer-aided diagnostic approach should be developed to reduce the
cost of brain MRI identification.Several efforts have been made to develop efficient and
reliable approaches for automatically categorizing brain tumors. These automated ap-
proaches mainly implemented by computer-aided medical image processing techniques
are increasingly aiding physicians in detecting brain tumors.

In recent years, Deep Learning (DL) has developed as a transformative technology in

medical imaging especially for tasks including diagnosis and classification. DL a subset
of Machine Learning, leverages deep neural models with multiple hidden layers to auto-
matically learn hierarchical patterns in data. Among DL models Convolutional Neural
Networks (CNNs) have shown their remarkable performance in analyzing MRI scans.
Their ability to automatically extract spatial features from raw images without the need
for handcrafted feature building makes them highly successful for clinical imaging tasks.
CNN-based models are increasingly surpassing traditional diagnostic methods in terms
of accuracy, reliability and efficiency. However training deep CNN models typically
requires large labelled datasets, which are often difficult to obtain in the medical do-
main due to privacy concerns, cost and annotation complexity. To address this challenge
transfer learning has become a widely adopted solution wherein pre-trained CNN such as
VGG16, ResNet50, DenseNet201, MobileNetV3Large and ConvNextBase are fine tuned
on smaller medical datasets to achieve high performance.
Chapter 1. Introduction 3

Despite their strengths, standard CNNs may not always focus on the most relevant
regions of medical images which can lead to suboptimal predictions. To overcome this
limitation, we propose the integration of a Dual Attention Residual Enhanced Network
(DARESNET) into these CNN architectures. The DARESNET module enhances each
model ability to selectively focus on the most informative spatial and channel-wise fea-
tures within MRI scans by improving diagnostic accuracy.The proposed approach has
been rigorously evaluated in two distinct classification settings: (1) binary classification
for distinguishing between healthy and tumor-affected brain scans and (2) four-class clas-
sification involving meningioma, pituitary, glioma tumors, and no-tumor cases. These
evaluations were conducted using three different publicly available MRI brain tumor
datasets to ensure robustness and generalizability across data sources. The core objec-
tives of this research are to improve classification performance, minimize false positives
and maintain computational efficiency for potential real-time clinical use. Through ex-
tensive quantitative experiments the study demonstrates the consistent effectiveness of
the DARESNET-enhanced models across all backbone networks and classification tasks.
Our key contributions of this paper are outlined as follows:
1) We introduce a Dual Attention Residual Network (DARESNET) module that in-
corporates both spatial and channel attention mechanisms. This enhances the models
ability to focus on the most informative features in brain MRI scans leading to more
accurate and robust tumor classification.
2) DARESNET is seamlessly integrated into several state-of-the-art CNN backbones like
MobileNetV3-Large, ConvNextBase, ResNet50, DenseNet201 and VGG16 demonstrat-
ing its flexibility and adaptability across diverse deep learning models.
3) The proposed models are evaluated under two distinct classification tasks to compre-
hensively assess their effectiveness. Firstly a binary classification task is performed to
distinguish between healthy and tumor-affected brain MRI scans, providing a founda-
tional assessment of the model’s ability to detect abnormality. Second, a more complex
four-class classification is conducted to identify and differentiate among meningioma,
pituitary, glioma and no tumor cases. This multi-class setup tests the models capability
to distinguish the presence of a tumor, also precisely categorize its type which is crucial
for supporting clinical decision-making and treatment planning.
4) The models are trained and tested on three publicly available brain MRI datasets,
ensuring robustness and generalizability of our results.
5) Experimental results show that our DARESNET-enhanced models consistently out-
perform existing state-of-the-art methods in terms of classification accuracy and other
performance metrics on the same datasets, establishing a new benchmark for brain tu-
mor classification.
Chapter 2

Literature Survey

2.1 Review of Binary Class Brain Tumor Classification

A. Hekmat et al. [1] presents an inventive show called the Attention-Fused MobileNet-
LSTM, outlined to make strides the location of brain tumors utilizing MRI. By combining
highlights from two set up CNN models,MobileNetV1 and MobileNetV2, this shows
challenges like high processing demands and noise in data. It upgrades include extraction
with transformers and employments gamma redress to capture imperative long-range
connections in images. Assessing its viability on a dataset of 3,000 MRI images, the
show accomplished an noteworthy accuracy of 98.66 %percent.With visualizations by
means of Grad-CAM, the model appears to be a “solid tool for medical diagnosis.

V Sukhavasi et al. [2] employs a pre-trained ResNet50 show for the binary classification
of brain tumors from MRI images. It includes pre-processing methods such as resizing,
normalization, and augmentation of the images.The ResNet50 model accomplished an
accuracy of 91.5%, outperforming conventional calculations like K-Nearest Neighbors
(87%), Support Vector Machine (88%), and Logistic Regression (69%).The discoveries
propose that ResNet50 altogether improves the accuracy of brain tumor discovery, which
can progress clinical decision-making and quiet results. Future research may investigate
its application in diagnosing other types of cancer.

R. Preetha et al. [3] presents a approach for automated brain tumor detection utilizing
MRI images, leveraging a refined EfficientNet-B4 CNN. Moreover model achieved an
accuracy of 99.33% on the Brain Tumor detection 2020 Kaggle dataset, beating other
deep learning models like VGG19 and ResNet. It also ensured model generalizability
through extensive feature extraction, hyperparameter optimization and K-Fold cross-
validation, the research guarantees the model’s generalizability. The approach highlights

4
Chapter 2. Literature Survey 5

the potential of deep learning in enhancing clinical decision-making for brain tumor
detection, advertising a quicker, more exact elective to manual diagnosis. Future research
may investigate transformer-based models and amplify this strategy to other therapeutic
imaging modalities.

Wageh M. et al. [4] proposes a Deep Features Concatenation and Machine Learning
Classifiers With Genetic Selection, combining deep learning and machine learning. Four
pre-trained CNN models (VGG-16, Initiation V3, ResNet-101, DenseNet-201) which are
concatenated and optimized for employing a hereditary calculation for feature selection.
The chosen highlights are classified utilizing SVM, Random Forest, Decision Tree, and
XGBoost. Trying on two datasets, the strategy achieved an accuracy of 99.74% and
99.87%, outperforming existing approaches. The integration of highlight concatenation
and hereditary determination improves execution, advertising a tool for early tumor
discovery. Future research incorporates extending datasets, refining tumor classification,
and exploring parallel processing to reduce computational time.

Sohaib Asif et al. [5] centers on Effectiveness of Different Deep Transfer Learning-
Based Models for Detecting Brain Tumors From MRI Images utilizing models like
Xception, NasNetLarge, DenseNet121 and InceptionResNetV2. The analysts prepro-
cessed MRI datasets to improve demonstrate execution, testing them with optimizers
like ADAM, SGD, and RMSprop. The Xception model, combined with the ADAM op-
timizer, achieved an accuracy of 99.67% on a larger dataset and 91.94% on a smaller
one. The results illustrate the viability of transfer learning in medical imaging, offering
a promising tool for early and exact brain tumor diagnosis, possibly improves patient
outcome.

2.2 Review of Multi-Class Brain Tumor Classification

Abdelmgeid A. Ali et al. [6] proposes a Co Evolutionary Genetic Algorithm (CEGA)

approach.By using pre-trained models EfficientNetB3 and DenseNet121,the study high-
lights the challenges like hyperparameter tuning and the need for larger datasets.CEGA
improves model accuracy and convergence speed, empowering successful classification
of gliomas, meningiomas, pituitary and no tumors.The technique effectively highlights
features from MRI , whereas attaining classification accuracy (99.39% for EfficientNetB3
and 99.01% for DenseNet121) without data augmentation, showing a significant progress
in computing brain tumor diagnosis and classification capabilities.

Nadia Bibi et al. [7] proposes a transfer learning-based approach utilizing the Incep-
tionV4 neural network to classify brain tumors from MRI images. The model prepared
Chapter 2. Literature Survey 6

on a dataset of 7022 MRI images, achieved an accuracy 98.7%, beating existing strate-
gies. The InceptionV4 known for its capacity to capture complex features, was refined to
handle the small details. The research about the significance of early and precise diag-
nosis, which can altogether move forward understanding results. Future research points
to improve model interpretability and coordinated multi-modal imaging data, possibly
revolutionizing brain tumor identification and treatment planning.

Ayesha Younis et al. [8] considers focus on improving the classification of irregular brain
tumors utilizing MRI images leveraging deep learning techniques. The proposed model
was based on a fine-tuned ResNet50, achieved an accuracy, with 99% on training and val-
idation datasets. It increased expectations in classifying glioma (98.33%), meningioma
(94.44%), no tumor (100%), and pituitary tumors (100%). The model utilized data
augmentation, regularization, and dropout layers to enhance performance, illustrating
accuracy, recall and F1 scores. Inspite its success, the model struggles with detecting
multiple tumors on a single scan, highlighting a future research about course. This work
highlights the potential of deep learning in advancing brain tumor detection.

Santosh Kumar Chhotray et al. [9] presents an optimized approach for brain tumor
classification utilizing MRI , leveraging pre-trained models like VGG16, ResNet50, NAS-
Net and DenseNet121, combined with consideration instruments such as CBAM. The
proposed Custom-CNN models, improved by the Bald Eagle Optimization (BEO) cal-
culation, beats conventional models in accuracy, sensitivity and precision over numerous
datasets (Br35H, FIGSHARE, SARTAJ). The integration of consideration components
highlights feature extraction,focuses on critical features within the images, whereas BEO
optimizes hyperparameters for quicker joining. This approach offers a strong, proficient
arrangement for medical imaging diagnostics, helping in early and exact brain tumor
discovery.

S Gopal Krishna Patro et al. [10] proposes a deep learning system for brain tumor clas-
sification utilizing MRI images, accomplishing 99.43% accuracy. By combining seven
pre-trained models (VGG-19, ResNet50V2, GoogLeNet, SqueezeNet, InceptionV3, Mo-
bileNetV2, and Xception), the approach minimizes false positives and improves models
quality. The dataset, sourced from Kaggle, incorporates 7023 images over four tumor
types. Preprocessing techniques like Gaussian blur and thresholding improvement in
image quality. Then ensemble method leverages each model’s qualities, outperforming
other models in accuracy , recall and F1-score. This advancement in medical imaging
offers a strong instrument for early tumor discovery, with future work focusing on huge
datasets and real-time applications for clinical use.

Lincy Annet Abraham et al. [11] proposes an progressed strategy, DC-YOLOv8FEN

for identifying brain tumors in MRI scans utilizing deep learning. By enhancing the
Chapter 2. Literature Survey 7

YOLOv8 by integrating dilated convolutions, self-shuffle consideration, and a dual fea-

ture pyramid network, which improves feature extraction and tumor classification ac-
curacy. The strategy achieves with 95.2% precision, 90.5% recall, 99.5% accuracy and
a 96.5% F1-score, outperforming existing methods. This advancement may altogether
help radiologists in real-time discovery, reducing delays and moving forward with quiet
results. Future work points to optimize the demonstrate for low-resource situations and
extend its clinical applications for way better brain tumor classification.

Adnan Saeed et al. [12] presents a dual branch ensemble method integrated with Gated
Global Local Consideration (GGLA) . Combining EfficientNetV2S and ConvNeXt, the
model captures both global and local tumor features, progressing classification accuracy.
A novel denoising technique improves image resolution, whereas GAN-based data aug-
mentation boosts model strength. Assessed on three class and four class datasets, the
model achieves 99.06% and 99.62% accuracy, beating existing strategies. The approach
addresses challenges like intra-class variation and inter-class similarity, advertising a
promising tool for precise brain tumor detection, possibly helping radiologists in clinical
decision-making.

Ji Hyeon Lee et al. [13] proposes improved computer aided diagnosis (CAD) framework
for classifying brain tumors in MRI looks utilizing deep learning. By applying Gaussian
channels for noise reduction and presenting Patterned-GridMask,a novel data augmen-
tation method, the framework makes strides the generalization and strength of profound
learning models. The approach accomplished up to 6% performance change over four
models, with the most noteworthy accuracy and F1 score coming to 97.74% and 97.75%,
individually. This progression underpins more precise early discovery, way better under-
standing results, and fitting treatment determination, tending to the challenges of brain
tumor determination and lessening misdiagnosis rates.

Mingyang Xu, Limei Guo, Hsiao-Chun Wu et al. [14] presents a novel, strong strategy
for programmed brain-tumor discovery and division utilizing MRI scans. The approach
combines dynamic form models with a unused texture-based examination to precisely
recognize and section tumors, indeed in complex cases with different tumors. The strat-
egy improves image quality, evacuates commotion, and disposes of untrue alerts by
sifting out little, non-tumor locales. Tried on a well-known brain-tumor dataset, the
method accomplished an AUC of 0.9244 and aan average correct segmentation area ra-
tio of 0.8019, outflanking existing strategies. This progression guarantees quicker, more
precise brain-tumor determination, diminishing human mistake and progressing quiet
results.
Chapter 2. Literature Survey 8

Anindya Nag et al. [15] presents TumorGANet, a deep learning model for classifying
brain tumors utilizing MRI scans. By combining ResNet-50 to highlight feature extrac-
tion and Generative Adversarial Networks (GANs) for data augmentation, the model
enhances dataset diversity and classification accuracy. The proposed system achieved
99.53% accuracy, with 100% precision and recall, and a 0.2% Hamming loss. More-
over, Explainable AI (LIME) is utilized for straightforwardness in features, which is
highlighted in the images. TumorGANet outperformed existing strategies, advertising
a tool for early and precise brain tumor discovery, possibly moving forward persistent
results and treatment planning. Future research aims to integrate federated learning
and advanced twin innovation.

Saravanan Srinivasan et al. [16] proposed a novel Brain-Tumor Detection Network

(BTDN). The method employs Contrast Limited Adaptive Histogram Equalization (CLAHE)
for improved image quality and data augmentation for robustness. The BTDN model
that was built to meet accuracy requirements for the tumor type discrimination is com-
pared with six deep learning systems. Furthermore, Secure-Net (SN) guarantees data
accuracy and secure transmission. The model’s accuracies are stunning given that it
has been validated against three different MRI datasets, demonstrating remarkable ac-
curacy. This method is highly useful in clinical practice because it integrates advanced
level deep learning with secure data handling in an efficient manner.

M. Azhagiri et al. [17] proposed an Enhanced AlexNet (EAN) model for enhanced
brain tumor classification. The approach starts with a preprocessing phase, in which
data augmentation is performed to improve the model’s generalization and robustness.
The EAN model adds augmented layers to the AlexNet architecture to optimize feature
extraction and classification accuracy. The system utilizes deep learning algorithms,
in which MRI images are automatically processed by the system with minimal human
involvement and improved diagnostic accuracy. The performance of the model is assessed
using accuracy, F1-score, precision, and recall, with high accuracy and low error. It
shows the efficiency of AI in precise brain tumor classification.

Mohamed Musthafa M et al. [18] proposed an improved method of brain tumor recogni-
tion by combining ResNet50 with Gradient-weighted Class Activation Mapping (Grad-
CAM) to increase both accuracy and understanding. Their approach starts with pre-
processing of MRI images and includes data augmentation for enhanced generalization
of the model. Deep learning modeling was performed using ResNet50 network as the
basic architecture and Grad-CAM visualized explanations of the model’s predictions to
make the decision easier to understand. The model is evaluated quantitatively in terms
Chapter 2. Literature Survey 9

of accuracy, precision, and recall, obtaining values above 98%. This research demon-
strates the relevance of explainable artificial intelligence in medical imaging diagnostics
by providing trust and clinical usability.

Suleyman A. AlShowarah et al. [19] introduced a deep learning method for the diagnosis
of brain tumors from MRI images. Extraction of features from the developed datasets
was made easy through the combination of Convolutional Neural Networks (CNN) and
VGG-19 as well as AlexNet. Random Forest (RF), Support Vector Machine (SVM), and
K-Nearest Neighbors (KNN) classification methods yielded the best classifier. For the
model to be reliable, its performance was evaluated on three training and testing data
splits. Merging of the data with KNN and SVM yielded results with 99.1% accuracy,
which confirms the efficiency of statistical and deep learning methods in tumor diagnosis.

Dr. Mahalakshmi et al. [20] applied fine-tuning through transfer learning using Effi-
cientNets to classify brain tumors. The technique entailed initiating EfficientNetB0 to
EfficientNetB4 using pre-trained weights of the ImageNet model and further fine-tuning
through additional top layers and one fully connected layer. Normalization as well as
data augmentation was done on the CE-MRI Figshare dataset with the aim to utilize
model generalizability. Training was performed with the Adam optimizer and categor-
ical cross-entropy loss. Model performance was measured based on accuracy, precision,
recall, and F1-score. Grad-CAM visualization was employed for model explanation
predictions, successfully identifying tumor areas in MRI scans for enhanced diagnostic
dependability.

2.3 Research Gaps

In spite of significant advancements in brain tumor detection utilizing deep learning, a

research gaps exist. Numerous studies considers achieves high accuracy on particular
datasets, but their generalizability to diverse, real-world datasets remains uncertain.
Existing models struggles to adjust in different image quality, resolution and tumor
types. Besides, most approaches focus on binary or multi-class classification, lacking
robust methodologies for identifying different tumors inside a single image.

Additionally, many authors have used pretrained CNN architectures for brain tumor
detection. However these model have might overfit due to limited range of dataset, also
it requires high computational resources and face generalization challenges.

The Dual Attention Residual Network (DARESNET) significantly enhances brain tumor
detection using MRI scans. By integrating Channel attention, Spatial attention and
Residual connections, DARESNET improves feature extraction, classification accuracy
Chapter 2. Literature Survey 10

and model stability. Channel attention focuses on the feature maps by assigning different
weights to each channel, by prioritizing data with decrease in noise. Spatial attention, on
the other hand, makes a difference as it allows the model to concentrate on spatial areas
inside the image, such as tumor affected areas, improving tumor detection. Residual
Connections solves the vanishing gradient problem, to maintain a better gradient flow
and stable training, which permits for more deep designs without performance loss.
By leveraging these components a precise tumor identification and strong generalization
over MRI datasets can be achieved. DARESNET improves both efficiency and accuracy,
making deep learning models more robust and clinically significant for real-world brain
tumor diagnosis.
Chapter 3

Methodology

The proposed deep learning model processes input images from dataset through prepro-
cessing and augumentation before feature extraction are done using a pretrained CNN
backbone. Preprocessing includes resizing, normalization and augmentation includes
flipping, brightness adjustments, saturation and contrast to improve efficiency, whereas
batching and prefetching optimize training speed. Also the model integrates a Dual
Attention Residual Network (DARESNET), which improves feature selection through
channel attention which enhances channel features and spatial to highlight key spatial
regions. These attention maps are combined with original features, preserving critical
details while refining discriminative data. Post Attention the global average pooling
reduces dimensions for feature integration. A softmax layer at that point classifies out-
puts based on numerous categories. The model benefits from transfer learning, dual
attention for improved feature selection and residual connections to avoid vanishing
gradients. This architecture is perfect for fine-grained image classification, offering a
balance between efficiency and accuracy by combining CNNs with lightweight attention
mechanisms as shown in Figure 3.1.

3.1 Database Used

3.1.1 Dataset-1

We have formulated dataset-1 of MRI images for brain tumor detection, compiled from
numerous sources, including open datasets from Kaggle, Roboflow and Mendeley. The
dataset comprises MRI images categorized into two classes: Tumor (3000 images) and
Healthy (2000 images). The MRI images strengths the soft tissue details, empowering

11
Chapter 3. Methodology 12

Figure 3.1: Methodology

a more exact investigation of brain tumors. Also, dataset is splitted in ratio of 80:10:10
for training, testing validation of images shown in Table 3,1.

Table 3.1: Dataset Distribution of Dataset-1

Class Training Testing Validation Total

Healthy 1600 200 200 2000
Tumor 2400 300 300 3000
Total 4000 500 500 5000
Chapter 3. Methodology 13

The model was trained and evaluated on multiple architectures. The performance met-
rics, including accuracy, precision, recall and F1 score were analyzed for each architec-
ture.

3.1.2 Dataset-2

The Dataset was obtained from the Kaggle website Br35H Brain Tumor Detection. It
consist of 3000 images, from that 1500 tumor images, while the remaining 1500 images
are non-tumor. Splitting data are mandatory for categorizing images.So the 1200 images
from non tumor class were used for training set,similarly same numbers are used for
tumor class as well. The testing subset were evaluated with 150 images from the non
tumor class and 150 images from the tumor class. Validation subset used 150 images
from the non tumor class and 150 images from the tumor class shown in Table 3.2.

Table 3.2: Dataset Distribution of Dataset-2

Class Training Testing Validation Total

Yes 1200 150 150 1500
No 1200 150 150 1500
Total 2400 300 300 3000

3.1.3 Dataset-3

The Dataset was obtained from the Kaggle website which includes 3 brain tumor datasets
from figshare, sartej and BR35H Brain Tumor Detection. It includes 7023 images, split-
ting data are mandatory for categorizing images therefore, so the dataset was splitted
into training,testing and validation subsets. For the training set we used 5617 images.
Similarly for the testing subset, we allocated 703 images from the all classes, also we
used 703 images from the all class to validation subset shown in Table 3.3.

Table 3.3: Dataset Distribution of Dataset-3

Class Training Testing Validation Total

Glioma 1297 162 162 1621
Meningioma 1315 165 165 1645
No Tumor 1600 200 200 2000
Pituitary 1405 176 176 1757
Total 5617 703 703 7023
Chapter 3. Methodology 14

3.2 Image Preprocessing

3.2.1 Creating Dataframe

A Dataframe is made to structure and manage the dataset images efficiently. This
initializes a dictionary storing image paths and their corresponding labels as shown in
Figure 3.2.This Dataframe serves as an input for preprocessing, normalization, and
augmentation, to maintain consistent integration with TensorFlow pipelines for deep
learning-based brain tumor classification.

Figure 3.2: Dataframe Mapping

3.2.2 Convert DataFrame into TensorFlow Dataset

The Dataframe is converted into a TensorFlow Dataset.This conversion takes image

paths and names from the dataframe and makes a dataset pipeline.Each image is stacked
and handled employing a preprocessing work, which reads the image file, decodes it,
resizes it, and normalizes pixel values.This TensorFlow dataset proficiently adds images
into the CNN model, guaranteeing smooth and scalable deep learning model shown in
Figure 3.3.

Figure 3.3: Tensorflow Dataset

Chapter 3. Methodology 15

3.2.3 Image Resizing

Image resizing could be a vital role in preprocessing step in deep learning to guar-
antee all input images have a uniform shape before being fed into the model.Since
images in a dataset may have different resolutions, resizing them to a settled mea-
surement (224×224 pixels) ensures consistency shown in Figure 3.4. The resizing is
performed which scales the image whereas protecting spatial connections.Keeping up the
aspect ratio, is vital to avoid distortion, which can negatively impact feature extraction.

Figure 3.4: Image Resizing

3.2.4 Image Normalization

Image normalization is a preprocessing step that scales pixel values to a standard range
by improving training stability and convergence in deep learning models. Raw im-
ages regularly have pixel values between 0 to 255 which can cause instability in gra-
dient flow. Normalization guarantees that pixel values drop inside a characterized
run, such as [0,1] (by isolating by 255) or [-1,1]. This standardization makes a dif-
ference in huge weight updates and accelerates model convergence shown in Figure 3.5.

For CNN models like ResNet, normalization adjusts input data with the models pre-
trained feature distributions. This improves feature extraction, reduces computational
cost and and improves overall classification accuracy in tasks like brain tumor detection.

Figure 3.5: Image Normalization

Chapter 3. Methodology 16

3.3 Image Augmentation

3.3.1 Image Flipping

Image flipping is a data augmentation technique that enhances a models by presenting

variations within the dataset. It makes a difference in model to learn spatial invariance
and making it more generalizable to unseen images.

In TensorFlow, flipping is performed randomly mirrors the images along the vertical axis.
It makes a difference in dataset without collecting extra images and reduces overfitting
as depicted in Figure 3.6. Also it must be utilized cautiously in settings where left-right
orientation matters. Combined with other augmentation techniques flipping improves
model generalization and performance.

Figure 3.6: Image Flipping

3.3.2 Image Brightness Adjustment

Image brightness adjustments is a data augmentation technique which increase proce-

dure that adjusts image illumination, making the model to learn more variations accord-
ing to shifting lighting conditions. In TensorFlow brightness is balanced and the most ex-
treme delta changes in brightness shown in Figure 3.7.

This method is especially useful in MRI scans where they have different lighting inten-
sities due to variations in equipments. By randomly expanding or reducing brightness
the model learns to recognize features independent of illumination by improving gener-
alization. However excessive brightness changes can distort important features, so it is
essential to fine-tune the augmentation ranges.
Chapter 3. Methodology 17

Figure 3.7: Image Brightness Adjustment

3.3.3 Image Saturation

Saturation is a data augmentation method which alters the intensity of an image. We

choose a irregular intensity figure from the run [0.8, 1.2] and adjusts the images color
escalated in like manner. In the event that saturation factor ¿ 1, colors gotten to be
more distinctive, on the off chance that saturation factor¡ 1, colors show up more blunt.
This change is performed by changing over the image to the HSV color space, altering
the saturation channel and changing over it back to RGB. Clipping ensures pixel values
stay inside substantial ranges shown in Figure 3.8. Random saturation shows strength
by presenting color varieties, avoiding overfitting to particular color distributions. It
is especially valuable in computer vision errands like image classification and object
detection. By improving generalization, this method makes a difference models perform
way better on images with different lighting conditions and color properties.

Figure 3.8: Image Saturation

3.3.4 Image Contrast

Image contrast improves visibility by adjusting pixel intensity differences, vital for MRI-
based tumor detection. Random contrast augmentation in deep learning improves model
generalization by exposing it to changing lighting conditions. In TensorFlow, contrast
adjustments help recognize tumor boundaries. High contrast sharpens edges, whereas
low contrast flattens images, requiring balanced adjustments to avoid artifacts shown in
Figure 3.9. Combining contrast adjustments with brightness adjustment, flipping, and
Chapter 3. Methodology 18

normalization strengthens model robustness, improving tumor classification accuracy

and reliability in MRI imaging applications.

Figure 3.9: Image Contrast

3.4 Image Batch & Prefetching

Batching and prefetching are essential techniques in TensorFlow to optimize training

efficiency. Batching groups different images into a single group, decreasing memory over-
head and speeding up model training.

Instead of handling one image at a time, the model processes an entire batch driving
to quicker computations and more steady gradients. Prefetching loads the next group
when the current group is being handled, minimizing idle time and improving GPU uti-
lization. This allows consistent data pipeline execution and avoid bottlenecks. Together
batching and prefetching optimize brain tumor detection by guaranteeing smooth and
high-performance preparing particularly when taking care of huge MRI datasets.

3.5 Pretrained CNN Backbone Models

Pretrained models have developed as a principal approach in deep learning for medical
image classification, essentially progressing model efficiency and performance as shown in
Figure 3.10. These models, at first prepared on large-scale datasets such as ImageNet,
capture fundamental visual features, decreasing the require for extensive labeled medical
data. Common pretrained models offers unique points of interest in terms of feature ex-
traction, computational efficiency and accuracy.

ResNet50 is a CNN architecture known for its residual connections which helps in solv-
ing vanishing gradients and improve feature propagation. It exceeds expectations in
learning features but can endure from overfitting when connected to small MRI datasets.
Chapter 3. Methodology 19

VGG16 a classic CNN with deep but uniform layers, is proficient in extracting texture-
based highlights, be that as it may, its high parameter count makes it computationally
costly compared to advanced options.

ConvNextBase a recent model inspired by Vision Transformers (ViTs), consolidates ad-

vanced design enhancements such as layer normalization and depth-wise convolutions,
empowering superior feature extraction for medical imaging.

DenseNet201 employments densely associated layers to enhance feature reuse, reduc-

ing redundancy and progressing gradient flow, which benefits fine-grained classification
tasks.

MobileNetV3Large designed for portable devices, utilizes depth-wise separable convolu-

tions to attain high efficiency with negligible computational cost, in spite of the fact that
it may need the capacity to capture important medical image features.

Figure 3.10: CNN Backbone

3.6 Dual Attention Residual Network

DARESNET enhances feature learning by integrating channel attention and spatial at-
tention within a residual framework. Channel attention emphasizes important feature
channels, whereas spatial attention highlights key spatial regions. The residual connec-
tion preserves original information, improving feature refinement and accuracy in image
classification.

3.6.1 Channel Attention

Channel Attention enhances deep learning models by focusing on the foremost impor-
tant feature channels whereas suppressing less relevant ones. It works by computing
Chapter 3. Methodology 20

global information from the feature map utilizing global average and global max pool-
ing. These two pooled features represent different contextual importance across channels.
Next, both pooled features pass through a shared multi layer perceptron with learnable
weights, followed by sigmoid activation to create attention scores for each channel. The
final channel attention map is obtained by element-wise summation of these scores.
This attention map is then multiplied with the significant channels of original feature
map, guaranteeing that the network learns which features matters most for the given
task as illustrated in Figure 3.11. By improving feature selection, Channel Atten-
tion enhances performance in tasks like MRI image classification. The mathematical
representation for channel attention are given from equations (3.1) to (3.6)

a) Global Average Pooling and Global Max Pooling:

H X
W
1 X
Favg = X(i, j, c) (3.1)
H ×W
i=1 j=1

Fmax = max X(i, j, c) (3.2)

i,j

where

• Favg is Average Pooled Feature Map

• Fmax is Max Pooled Feature Map

• X(i, j, c) is the feature map value at spatial position (i, j) for channel c, and H, W
are the height and width of the feature map

b) Fully Connected Layers & Activation:

Mavg = σ(W2 δ(W1 Favg )) (3.3)

Mmax = σ(W2 δ(W1 Fmax )) (3.4)

where:
Chapter 3. Methodology 21

• Mavg & Mmax are Attention Weight Vectors or Modulation Factors

• W1 , W2 are learnable weights,

• δ is the ReLU activation,

• σ is the sigmoid activation.

c) Final Channel Attention Weights:

Mc = Mavg + Mmax (3.5)

d) Refined Feature Map:

X ′ = Mc · X (3.6)

where · represents channel-wise multiplication.

Figure 3.11: Channel Attention Module

3.6.2 Spatial Attention

Spatial Attention focus on important spatial regions in an image whereas suppressing

less relevant regions. Unlike channel attention which focus on important channels, spa-
tial attention highlights key spatial regions where essential features are present. It begins
by applying max pooling and average pooling along the channel dimension, making two
feature maps that capture global and local spatial context. These two maps are added
and passed through a 7×7 convolution layer to learn spatial dependencies. The output is
then activated using sigmoid function, creating a spatial attention map. This attention
map is multiplied element-wise with the original feature map, improving critical spatial
regions whereas filtering out noise. Spatial Attention is especially useful in object de-
tection and MRI imaging, where identifying key regions is vital. By improving spatial
feature selection, it helps networks learn where to focus, driving to better predictions and
improved performance as illustrated in Figure 3.12. The mathematical representation
for spatial attention are given from equations (3.7) to (3.10).
Chapter 3. Methodology 22

a) Spatial Pooling (Average & Max):

C
1 X
Favg (i, j) = X(i, j, c) (3.7)
C
c=1

Fmax (i, j) = max X(i, j, c) (3.8)

c

where C is the number of channels, and (i, j) represents the spatial position.

b) Concatenation & Convolution:

Fspatial = σ(f7 ([Favg , Fmax ])) (3.9)

where:

• Fspatial represents a spatial attention map

• f7 is a 7 × 7 convolution,

• σ is the sigmoid activation function.

c) Final Spatial Attention Weights:

X ′′ = Fspatial · X ′ (3.10)

where · represents element-wise multiplication.

Figure 3.12: Spatial Attention Module

3.6.3 Residual Connections

Residual Connections allow neural networks to skip layers and pass information directly
from one layer to another. This prevents vanishing gradients and makes training deep
Chapter 3. Methodology 23

models more stable and efficient shown in Figure 3.13 . Instead of learning the full
transformation, the network learns the residual (difference) between input and output,
improving gradient flow.

In DARESNET, Residual Connections help preserve important features, allow for deeper
architectures, and enhance tumor detection in MRI scans by improving feature extrac-
tion and classification accuracy. Mathematically, it follows equation (3.11).

Xres = X + X ′ + X ′′ (3.11)

where:

• X is the input feature map,

• X ′ is the channel-refined output,

• X ′′ is the spatial-refined output.

Figure 3.13: Residual Connections

3.7 GlobalAveragePooling2D

Global Average Pooling2D reduces the spatial dimensions of feature maps by computing
the average values of each channel. Rather than smoothing the whole feature outline,
it condenses each channel into a single scalar, protecting spatial data in a compact
form.For example, an input highlight outline encompasses a shape (H, W, C), global
average pooling outputs a (1, 1, C) vector. In models like MobileNet, global average
pooling minimizes parameters, reducing overfitting whereas keeping up important fea-
tures. It also changes feature maps into a 960-dimensional vector, making it effective
for classification into dense layers.
Chapter 3. Methodology 24

3.8 Dense (128, ReLU Activation)

A Dense layer may be a completely associated neural network layer with 128 neurons
and rectified linear unit activation. Each neuron learns designs from input features by
applying weights, biases, and activations. The ReLU activation presents non-linearity,
also to show complex connections. This layer refines learned features, improves model
performance for classification. Also it is regularly utilized after feature extraction, it
makes a difference in dimensionality reduction and decision-making. The 128 neurons
act as an halfway representation for last classification. It also adds non-linearity and
extracts refined patterns.

3.9 Dense (n, Softmax Activation)

A Dense layer has n neurons, classified into n categories. The Softmax work changes
over logits into likelihood scores summing to 1, empowering classification. The model
predicts the image, making it perfect for classification with probabilistic certainty.

3.10 Output

It represents the final classification result, determined from the predicted probabilities.
The model assigns a class label based on the highest probability from the softmax output.
This decision chooses the final diagnosis, significant for MRI image diagnosis and tumor
detection systems.
Chapter 4

Results and Discussions

Objective

The objective of this study is to enhance the performance of deep learning-based im-
age classification using DARESNET, an optimized neural network architecture. This
involves improving model accuracy, efficiency, and robustness through strategic dataset
preprocessing, hyperparameter tuning, and phase-wise model training. The study fo-
cuses on optimizing computational resources while maintaining high classification preci-
sion, making the model suitable for real-world applications. Additionally, a comparative
analysis with existing architectures evaluates the proposed approach in terms of accu-
racy, convergence speed and generalization. The ultimate goal is to develop a scalable,
high-performance model that can be deployed effectively for larger image recognition
tasks.

The results presented in this study evaluate the performance of the proposed DARES-
NET model in terms of classification accuracy, computational efficiency, and generaliza-
tion capability. Comparative analysis is conducted against existing models to demon-
strate improvements achieved through optimized dataset preprocessing, hyperparameter
tuning, and phase-wise training.

4.1 Dataset Splitting & Hyperparameters Tuning

Each dataset was initially divided into a 80:10:10 split for training, testing and validat-
ing. Of the 80% allocated for training, 10% for testing and 10% validation set. This
validation set was used exclusively for tuning hyperparameters and monitoring model

25
Chapter 4. Proof of Implementation and Result comparison 26

performance during training, ensuring that the test and validation set remained com-
pletely independent for final evaluation. Additionally, Table 4.1 presents the various
parameters utilized during the experimentation with the models and the optimization
algorithm.
Table 4.1: Optimized Hyperparameters

Parameter Optimized Value

Image Size 224
Batch Size 128
Epochs 10
Learning rate 0.1
Execution Environment GPU
Contrast Adjustments 0.8-1.2
Brightness Adjustments 0.8-1.2
Dense Units 128
Loss Function Binary & Categorical cross-entropy

4.2 Model Evaluation

After training the DARESNET model using the optimized hyperparameters, the next
step is to assess its performance on the test dataset and accurately classify brain tumors
into their respective categories. The model’s predictions on the test set are compared
with ground truth labels to compute key evaluation metrics, including accuracy, preci-
sion, recall, and F1-score.

4.2.1 Accuracy and Loss Plots

Accuracy and loss plots help visualize model performance during training and validation.
These plots guide us in diagnosing underfitting, overfitting, and convergence issues. Loss
is a measure of how well or poorly a model performs on a given dataset. It represents the
error between the predicted and actual values. Lower loss indicates a better-performing
model. Accuracy measures the proportion of correct predictions. Higher Accuracy
indicates a better-performing model.

4.2.2 Confusion Matrix

A Confusion Matrix evaluates classification models by comparing true and predicted

labels. It comprises of True Positives (TP), True Negatives (TN), False Positives (FP),
Chapter 4. Proof of Implementation and Result comparison 27

and False Negatives (FN), as shown in Figure 4.1 Key metrics derived from the con-
fusion matrix include Accuracy (overall correctness), Precision (proportion of correctly
predicted positives), Recall (true positive rate), and F1-score (harmonic mean of preci-
sion and recall). Additionally, Specificity measures the true negative rate. This matrix
helps in identifying issues such as overfitting, class imbalance, and model errors. While
high recall is crucial for applications like disease detection, high precision is essential for
fraud detection. By analyzing misclassification patterns, the Confusion Matrix provides
valuable insights to enhance model performance.

Figure 4.1: Training Accuracy Plot

4.2.3 Precision

Precision measures the proportion of correctly predicted positive instances out of all
instances classified as positive. It reflects the reliability of the model when making
positive predictions.
TP
Precision =
TP + FP

4.2.4 Recall

Recall measures the proportion of correctly predicted positive instances out of all actual
positive cases. It indicates the model’s effectiveness in identifying positive cases.

TP
Recall =
TP + FN

4.2.5 F1-Score

F1-score is the harmonic mean of precision and recall, offering a balanced evaluation of
both metrics. It is particularly useful when false positives and false negatives hold equal
significance.
2 × Precision × Recall
F1 Score =
Precision + Recall
Chapter 4. Proof of Implementation and Result comparison 28

4.2.6 Confidence Scores

A confidence score represents the model’s certainty in its prediction, typically ranging
from 0 to 1 or expressed as a percentage. A higher score indicates strong confidence in
the prediction, while a lower score suggests uncertainty. It is derived from probability
distributions across classes, where the highest probability determines the predicted label.
Unlike accuracy, which evaluates overall correctness, confidence is specific to individual
predictions.

4.3 Comparison of CNN-DARESNET VS CNN Architec-

tures using Dataset 1

The Dataset 1 which we used in this phase is described above in database used where
80% of the data were randomly picked, also used to train the proposed model and 20%
of the data were used to evaluate the model by testing and validating. The training
accuracy of this phase are shown in Figure 4.2.The proposed model was compared
with 5 different CNN architectures.The model attained the highest validation accuracy
of 99.40% , training accuracy of 99.73% and with testing accuracy 99%. The training
and validation loss of the proposed model were plotted, shown in Figure 4.4. The
implied proposed mode training and validation accuracy are shown in Figure 4.5.

The best models validation loss was observed to be 0.0084.The detailed confusion matrix
is illustrated in Figure 4.6. Also the classification report are presented in Table 4.2.
The confidence scores of the model which obtained the highest accuracy is shown in Fig-
ure 4.3. In this phase the accuracy of all the CNN-models were improved significantly
when these architecture were integrated with DARESNET.

Figure 4.2: Training Accuracy Plot

Chapter 4. Proof of Implementation and Result comparison 29

Table 4.2: Classification Report and (Testing & Validation ) Accuracy

Model Testing Accuracy Validation Accuracy Precision Recall F1 Score

MobileNetV3Large 98% 98.80% 0.97 0.97 0.97
DARES-MobileNetV3Large 99% 99.40% 0.98 0.98 0.99
ConvNextBase 99% 98.80% 0.99 0.98 0.98
DARES-ConvNextBase 99.20% 99.40% 0.98 0.98 0.98
ResNet-50 96.80% 97.60% 0.96 0.96 0.96
DARES-ResNet-50 97.80% 98.80% 0.97 0.96 0.97
Densenet201 92.80% 91.40% 0.92 0.92 0.91
DARES-DenseNet201 92.80% 93.60% 0.91 0.91 0.91
VGG-16 97.20% 96.60% 0.97 0.97 0.97
DARES-VGG16 97.20% 96.60% 0.98 0.98 0.97

Figure 4.3: Confidence Scores of the model which achieved highest accuracy
Chapter 4. Proof of Implementation and Result comparison 30

Figure 4.4: Training vs Validation Loss Plot

Chapter 4. Proof of Implementation and Result comparison 31

Figure 4.5: Training vs Validation Accuracy Plot

Chapter 4. Proof of Implementation and Result comparison 32

Figure 4.6: Confusion Matrix of the model

4.4 Classification Results of CNN-DARESNET using Dataset

2

The Dataset 2 which we used in this phase is described above in database used. The
training accuracy are shown in Figure 4.7.The Proposed model was compared with
5 distinct CNN architectures integrated with DARESNET. The model attained the
maximum validation accuracy of 98.80% , training accuracy of 99.03% and with testing
accuracy 98% . The training and validation loss of model were plotted , shown in Figure
4.9. The proposed mode training and validation accuracy is shown in Figure 4.10. The
best model validation loss is 0.02. The detailed confusion matrix is depicted in Figure
4.11. The Classification Report of all the parameters are presented in shown in Table
4.3. Also the confidence scores of the model which obtained the highest accuracy is
shown in Figure 4.8.
Chapter 4. Proof of Implementation and Result comparison 33

Figure 4.7: Training Accuracy Plot

Table 4.3: Classification Report and (Testing Validation ) Accuracy

Model Testing Accuracy Validation Accuracy Precision Recall F1

MobileNetV3Large 98% 98.80% 0.97 0.97 0.97
ConvNextBase 98.67% 98.33% 0.98 0.98 0.98
ResNet-50 93% 94.67% 0.90 0.88 0.88
Densenet201 91.33% 90% 0.96 0.91 0.91
VGG-16 93.33% 95.67% 0.90 0.90 0.89

Figure 4.8: Confidence Scores of the model which achieved highest accuracy
Chapter 4. Proof of Implementation and Result comparison 34

Figure 4.9: Training vs Validation Loss Plot

Figure 4.10: Training vs Validation Accuracy Plot

Chapter 4. Proof of Implementation and Result comparison 35

Figure 4.11: Confusion Matrix of the model

4.5 Classification Result of CNN-DARESNET using Dataset

3

The Dataset 3 which we used in this phase is described above in database used. The
training accuracy of the Figure 4.12. The proposed model was compared with 5 dif-
ferent CNN architectures integrated with DARESNET.The model attained with highest
validation accuracy of 95.54% , training accuracy of 99.26% and testing accuracy 96.16%.
The training and validation loss training were also plotted, shown in Figure 4.14. The
proposed model training and validation accuracy is shown in Figure 4.15. The best
models validation loss is 0.0329. The detailed confusion matrix are depicted in Figure
4.16.The Classification Report of all the parameters are presented in shown in Table
4.4. Also the confidence scores of the model which obtained the highest accuracy is
shown in Figure 4.13.

Table 4.4: Classification Report and (Testing Validation ) Accuracy

Model Testing Accuracy Validation Accuracy Precision Recall F1

MobileNetV3Large 91.18% 92.45% 0.92 0.91 0.91
ConvNextBase 96.16% 95.44% 0.96 0.96 0.96
ResNet-50 93.17% 92.14% 0.93 0.92 0.92
Densenet201 91.75% 91.45% 0.91 0.91 0.91
VGG-16 92.75% 95.44% 0.92 0.92 0.92
Chapter 4. Proof of Implementation and Result comparison 36

Figure 4.12: Training Accuracy Plot

Figure 4.13: Confidence Scores of the model which achieved highest accuracy
Chapter 4. Proof of Implementation and Result comparison 37

Figure 4.14: Training vs Validation Loss Plot

Figure 4.15: Training vs Validation Accuracy Plot

Chapter 4. Proof of Implementation and Result comparison 38

Figure 4.16: Confusion Matrix of the model

4.6 Comparison on Existing and Previous Study

Advancements in brain tumor classification have largely been driven by deep learning
and transfer learning, as demonstrated in recent studies. Our proposed method builds
upon these foundations while addressing the limitations observed in previous research.
By incorporating innovative techniques, our approach enhances accuracy and robust-
ness in tumor detection. A comprehensive comparison with existing methods highlights
the strengths of our model and demonstrates its advantages in effectively capturing
intricate image features, improving classification performance, and minimizing errors.
These improvements make our approach a valuable contribution to AI-driven medical
diagnostics.
Chapter 4. Proof of Implementation and Result comparison 39

4.6.1 Comparison with Binary Classification Models

Our method demonstrates superior performance compared to binary classification ap-

proaches by leveraging an optimized deep learning framework that integrates feature ex-
traction, data augmentation, and hyperparameter tuning. The Attention-Fused MobileNet-
LSTM achieves the highest accuracy; however, while their approach utilizes transform-
ers for feature extraction, it struggles with high computational cost. Our model effec-
tively reduces computational overhead while maintaining high accuracy by incorporating
ResNet50 with advanced pre-processing techniques. While their strategy outperforms
traditional classifiers,

our approach surpasses it by implementing enhanced augmentation and feature selection

methods. Additionally, their fine-tuned EfficientNet-B4 model showed improvements
over VGG19 and ResNet50, but our method further enhances interpretability and feature
learning by integrating attention mechanisms. Another existing study introduced deep
feature concatenation with genetic selection. Although highly accurate, their approach
relies on computationally intensive classifiers, whereas our model optimizes feature ex-
traction while sustaining competitive accuracy. One study explored transfer learning
using Xception and NasNet, and our method builds upon their findings by improving
generalization through advanced augmentation and optimization techniques.

4.6.2 Comparison with Multi-Class Classification Models

In multi-class classification, our method delivers competitive performance while address-

ing key challenges related to dataset limitations. A previous study proposed a Co-
Evolutionary Genetic Algorithm (CEGA) integrated with EfficientNetB3 and DenseNet121.
While their approach excels in convergence speed, our model introduces novel regular-
ization techniques to enhance stability across diverse datasets. Another study utilized
InceptionV4, which, despite being effective, lacks interpretability. Our approach incorpo-
rates explainability techniques to improve decision transparency. Similarly, a fine-tuned
ResNet50 was employed, demonstrating high accuracy across various tumour types; how-
ever, our method enhances class differentiation between overlapping categories through
advanced feature selection mechanisms.

Another approach integrated CBAM attention components with pre-trained models and
optimized performance using Bald Eagle Optimization (BEO). While their method sig-
nificantly improves accuracy, our model refines feature extraction to reduce misclas-
sifications. Some studies have explored ensemble strategies, such as combining seven
pre-trained models. Although ensemble methods enhance reliability, they introduce re-
dundancy, whereas our technique optimizes computational efficiency while maintaining
Chapter 4. Proof of Implementation and Result comparison 40

high classification performance. A real-time detection model, DC-YOLOv8FEN, pri-

oritizes speed but compromises sensitivity. In contrast, our model balances real-time
applicability with robust classification accuracy.

Further, a GGLA-NeXtE2NET model combined EfficientNetV2S and ConvNext, effec-

tively capturing local and global tumour features, while our approach enhances fea-
ture extraction with advanced augmentation techniques. Another study improved deep
learning robustness using Patterned-Grid Mask, whereas our method expands augmen-
tation techniques to enhance generalization. A texture-based segmentation method im-
proved segmentation performance, but our model ensures superior classification accu-
racy through hybrid feature extraction. Finally, Tumor-GANet leveraged GAN-based
augmentation to enhance data diversity, whereas our approach applies balanced aug-
mentation strategies to prevent overfitting.

By integrating dual attention mechanisms, our model captures both spatial and channel-
wise information, leading to superior tumour localization and classification accuracy.
Unlike most studies that focus solely on T1-weighted MRI, our method incorporates
multi-sequence MRI images, improving classification reliability. Additionally, our model
achieves high accuracy with fewer parameters compared to approaches relying on mul-
tiple pre-trained models.

By, integrating data augmentation and weighted loss functions, we effectively address
the issue of imbalanced datasets, a common challenge in brain tumour classification.
Through cross-validation and hyperparameter tuning, our model demonstrates superior
generalizability across different datasets, minimizing overfitting. Compared to existing
methods, our approach achieves an optimal balance between accuracy, computational
efficiency, and interpretability. By integrating advanced augmentation, feature selection,
and attention mechanisms, our model outperforms conventional deep learning models in
both binary and multi-class classification settings. Future research will focus on real-time
deployment and expanding the dataset to further validate the model’s generalizability.

We compared the accuracy of our proposed CNN-DARESNET model for binary and
multi-class classification with the accuracies obtained by other authors for same classi-
fication problem as shown in Tables 4.5, 4.6 & 4.7. From the tables we can clearly
observe that our proposed model performed better than the existing works on the similar
databases.
Chapter 4. Proof of Implementation and Result comparison 41

Table 4.5: Comparison of Brain Tumor Detection Methods

Author Database Used Methodology Accuracy

Mohamed
Brain Tumor MRI im-
Musthafa Utilized ResNet50 with
ages for Brain Tumor 98.30%
M et al. Grad-CAM
Detection
[16]

Combined dataset of
Developed a Brain Tu-
Sandeep Br35H, Brats, and Kag-
mor Detection Network
Kumar et gle Data Repository 98.81%
(BTDN) using deep
al. [17] which consists of Brain
learning
Tumor MRI images

Suleyman Deep learning models

A. Al- Br35H Brain Tumor with VGG-19 and
99.1%
Showarah Detection AlexNet for feature
et al. [18] extraction

Enhanced AlexNet
M. Azhagiri Brats2015 dataset from
model with data aug- 99.32%
et al. [19] Kaggle
mentation techniques

CE-MRI Figshare Fine-tuning Efficient-

Baiju Babu
dataset for brain tumor Net architecture with 99.06%
et al. [20]
classification transfer learning

Implemented Dual At-

Proposed tention Residual mecha-
Dataset -1 99.73%
Model nism & integrated with
MobileNetV3Large
Chapter 4. Proof of Implementation and Result comparison 42

Table 4.6: Comparison of review on binary class brain tumor classification using
Dataset-2

Author Methodology Accuracy

A Hekamat et al. Attention based CNN architecture (Mo- 98.66%

[1] bileNetV1 and MobileNetV2) with LSTM
networks

V. Sukhavasi et Comparison of Pre-trained ResNet50 91.5%

al. [2] model with traditional methods

R. Preetha et al. Development of a fine-tuned EfficientNet- 99.33%

[3] B4 model

M. Wageh et al. Genetic algorithm applied to select sig- 99.74%

[4] nificant features before classification us-
ing machine learning classifiers (SVM, RF,
DT, KNN)

S. Asif et al. [5] Transfer learning applied on pre-trained 99.67%

models for optimized classification

Proposed Model Implemented Dual Attention Resid- 99.03%

ual Network & integrated with Mo-
bileNetV3Large
Chapter 4. Proof of Implementation and Result comparison 43

Table 4.7: Comparison of review on multi-class brain tumor classification using

Dataset-3

Source Methodology Accuracy

Abdelmegid A. Ali et al. [6] Deep Learning Model with Co-Evolutionary 99.39%
Genetic Algorithm (CEGA) for optimizing hy-
perparameters

Nadia Bibi et al. [7] Transfer learning-based Inception V4 98.7%

Ayesha Younis et al. [8] Transfer learning-based ResNet50 model 99%

Santosh Kumar et al. [9] Usage of a Custom-CNN model which includes 97.22%
pre-trained models (VGG16, ResNet50, NAS-
Net, DenseNet121) and attention mechanisms
(CBAM).

S Gopal Krishna Patro et al. [10] Ensemble framework with seven deep learning 99.43%
models

Lincy Anter Abraham et al. [11] DC-YOLOv8FEN integrating self-shuffle at- 99.5%
tention, dual feature pyramid network, and di-
lated convolutions

Adnan Saeed et al. [12] GGLA-NeXtE2NET model combining Dual- 99.06%

Branch Ensemble network with a Gated
Global-Local Attention mechanism

Ji-Hyeon Lee et al. [13] Enhanced computer-aided diagnosis (CAD) al- 97.74%
gorithm incorporating Gaussian filters and a
novel Patterned-GridMask for detection of Tu-
mor

Mingyang Xu et al. [14] A ctive-contour model using an outward ex- 92.44

pansion approach for detection of tumor and
usage of texture-based decision metrics

Anindya Nag et al. [15] TumorGANet model by integrating ResNet50 99.53%

for feature extraction and GANs

Proposed Model Implemented Dual Attention Residual Net- 99.26%

work& integrated with ConvNextBase
Chapter 5

Conclusion and Future Scope

The proposed Double Attention Residual Network (DARESNET) model demonstrates

exceptional performance in brain tumor classification by capturing intricate image fea-
tures and outperforming widely used CNN models like ResNet50, EfficientNetB7, and
VGG16. Achieving high accuracy rates of 99.40%, 99.03%, and 99.26% with low loss val-
ues across three datasets, it proves robust in medical image analysis. The classification
report and confusion matrix confirm its effectiveness, with high precision, recall, and
F1-score, along with minimal false positives and false negatives, ensuring reliable tumor
detection. Unlike traditional CNN architectures, DARESNET utilizes double attention
mechanisms and residual learning to extract meaningful patterns while reducing overfit-
ting. The smooth convergence of training and validation loss curves further supports its
stability and learning efficiency. By integrating advanced deep learning techniques, this
study enhances AI-driven medical diagnostics, offering accurate, efficient, and scalable
solutions for clinical applications.

Future research should focus on integrating multi-modal imaging, such as combining

MRI and CT scans, to improve diagnostic accuracy. Explainable AI techniques can
enhance model interpretability, helping clinicians understand AI-driven decisions and
increasing trust in healthcare. Optimizing real-time tumor detection with edge com-
puting can enable faster, on-site analysis, reducing reliance on centralized systems and
making AI-powered diagnostics more accessible in remote areas. Further advancements
in self-supervised learning and data augmentation can improve adaptability to unseen
medical imaging datasets, while federated learning may allow collaborative AI training
across hospitals while preserving patient privacy. As deep learning evolves, AI applica-
tions in healthcare will enhance accuracy, speed, and accessibility in tumor detection,
ultimately benefiting both clinicians and patients.

44
Chapter 6

Appendix

import tensorflow as tf
from tensorflow.keras.layers import Input, Conv2D, GlobalAveragePooling2D, Global-
MaxPooling2D, Dense, Multiply, Add, Reshape
from tensorflow.keras.applications import MobileNetV3Large
from tensorflow.keras.models import Model

def channelattention(inputfeature, ratio=8):

channel = inputfeature.shape[-1]
Global Max Pooling
maxpool = GlobalMaxPooling2D()(inputf eature)
maxpool = Dense(channel/ratio, activation = relu)(maxpool)
maxpool = Dense(channel, activation = sigmoid)(maxpool)
GlobalAverageP ooling
avgpool = GlobalAverageP ooling2D()(inputf eature)
avgpool = Dense(channel/ratio, activation = relu)(avgpool)
avgpool = Dense(channel, activation = sigmoid)(avgpool)
M ergebothattentionscores
channelattentionf eature = Add()([maxpool, avgpool])
channelattentionf eature = Reshape((1, 1, channel))(channelattentionf eature)

return Multiply()([inputfeature, channelattentionfeature])

def spatialattention(inputfeature):
avgpool = tf.reducemean(inputfeature, axis=-1, keepdims=True)

45
Chapter 6. Appendix 46

maxpool = tf.reducemax(inputf eature, axis = −1, keepdims = T rue)

concat = tf.concat([avgpool, maxpool], axis = −1)

conv = Conv2D(filters=1, kernelsize=7, padding=same, activation=sigmoid)(concat)

return Multiply()([inputfeature, conv])

def dualattentionresidualblock(inputfeature):
channelrefined = channelattention(inputfeature)
spatialrefined = spatialattention(channelrefined)
residual = Add()([inputfeature, channelrefined,spatialrefined]) Residual connection
return residual

backbone = MobileNetV3Large(weights=imagenet, includetop=False, inputshape=inputshape)

backbone.trainable = False Freeze the pretrained backbone

inputlayer = Input(shape=inputshape)
x = backbone(inputlayer) Extract features from backbone

Apply Dual Attention Residual Block

x = dualattentionresidualblock(x)

Classification head
x = GlobalAveragePooling2D()(x)
x = Dense(128, activation=relu)(x)
outputlayer = Dense(numclasses, activation=softmax)(x)

model = Model(inputs=inputlayer, outputs=outputlayer)

model.summary()
Chapter 7

Bibliography

Hekmat, A., Zhang, Z., Khan, S. U. R., Shad, I., & Bilal, O. (2025). An attention-fused
architecture for brain tumor diagnosis. Biomedical Signal Processing and Control, 101,
107221.

Sukhavasi, V., Gangisetty, N., & Maharaj, N. S. (2023, December). Accuracy Prediction
for Detecting Brain Tumour from MRI Images using ResNet50. In 2023 3rd International
Conference on Innovative Mechanisms for Industry Applications (ICIMIA) (pp. 420-
426). IEEE.

Preetha, R., Priyadarsini, M. J. P., & Nisha, J. S. (2024). Automated Brain Tumor
Detection from Magnetic Resonance Images Using Fine-Tuned EfficientNet-B4 Convo-
lutional Neural Network. IEEE Access.

Wageh, M., Amin, K., Algarni, A. D., Hamad, A. M., & Ibrahim, M. (2024). Brain Tu-
mor Detection Based on Deep Features Concatenation and Machine Learning Classifiers
With Genetic Selection. IEEE Access.

Asif, S., Yi, W., Ain, Q. U., Hou, J., Yi, T., & Si, J. (2022). Improving effectiveness
of different deep transfer learning-based models for detecting brain tumors from MR
images. IEEE Access, 10, 34716-34730.

Ali, A. A., Hammad, M. T., & Hassan, H. S. (2025). A Co-Evolutionary Genetic

Algorithm Approach to Optimizing Deep Learning for Brain Tumor Classification. IEEE
Access.

Ali, A. A., Hammad, M. T., & Hassan, H. S. (2025). A Co-Evolutionary Genetic

Algorithm Approach to Optimizing Deep Learning for Brain Tumor Classification. IEEE
Access.
47
Chapter 7. Bibliography 48

Younis, A., Qiang, L., Afzal, Z., Adamu, M. J., Kawuwa, H. B., Hussain, F., & Hussain,
H. (2024). Abnormal brain tumors classification using resnet50 and its comprehensive
evaluation. IEEE Access.

Chhotray, S. K., Mishra, D., Pati, S. P., & Mishra, S. (2025). An Optimized Cascaded
CNN Approach for Feature Extraction from Brain MRIs for Tumor Classification. IEEE
Access.

Patro, S. G. K., Govil, N., Saxena, S., Mishra, B. K., Zamani, A. T., Miled, A. B., ... &
Hamdan, M. (2024). Brain Tumor Classification Using an Ensemble of Deep Learning
Techniques. IEEE Access.

Abraham, L. A., Palanisamy, G., & Goutham, V. (2025). Dilated Convolution and
YOLOv8 Feature Extraction Network: An Improved Method for MRI-based Brain Tu-
mor Detection. IEEE Access.

Saeed, A., Shehzad, K., Bhatti, S. S., Ahmed, S., & Azar, A. T. (2025). GGLA-
NeXtE2NET: A Dual-Branch Ensemble Network with Gated Global-Local Attention
for Enhanced Brain Tumor Recognition. IEEE Access.

Lee, J. H., Chae, J. W., & Cho, H. C. (2024). Improved classification of different brain
tumors in mri scans using patterned-gridmask. IEEE Access.

Xu, M., Guo, L., & Wu, H. C. (2024). Novel robust automatic brain-tumor detection
and segmentation using magnetic resonance imaging. IEEE Sensors Journal.

Nag, A., Mondal, H., Hassan, M. M., Al-Shehari, T., Kadrie, M., Al-Razgan, M., ...
& Bairagi, A. K. (2024). TumorGANet: A transfer learning and generative adversarial
network-based data augmentation model for brain tumor classification. IEEE Access.

AlShowarah, S. A. (2025). DeepCancer: deep learning for brain tumor detection-based

application system. Neural Computing and Applications, 1-20.

Azhagiri, M., & Rajesh, P. (2024). EAN: enhanced AlexNet deep learning model to de-
tect brain tumor using magnetic resonance images. Multimedia Tools and Applications,
83(25), 66925-66941.

T. R, M., V, V. K., & Guluwadi, S. (2024). Enhancing brain tumor detection in MRI
images through explainable AI using Grad-CAM with Resnet 50. BMC medical imaging,
24(1), 107.
Chapter 7. Bibliography 49

Mathivanan, S. K., Srinivasan, S., Koti, M. S., Kushwah, V. S., Joseph, R. B., & Shah,
M. A. (2025). A secure hybrid deep learning framework for brain tumor detection and
classification. Journal of Big Data, 12(1), 72.

Babu Vimala, B., Srinivasan, S., Mathivanan, S. K., Mahalakshmi, Jayagopal, P., &
Dalu, G. T. (2023). Detection and classification of brain tumor using hybrid deep
learning models. Scientific reports, 13(1), 23029.

Sachdeva, J., Sharma, D., & Ahuja, C. K. (2024). Comparative analysis of different deep
convolutional neural network architectures for classification of brain tumor on magnetic
resonance images. Archives of Computational Methods in Engineering, 31(4), 1959-1978.

Mehmood, Y., & Bajwa, U. I. (2024). Brain tumor grade classification using the Con-
vNext architecture. Digital Health, 10, 20552076241284920.

Ahmad, S., & Choudhury, P. K. (2022). On the performance of deep transfer learning
networks for brain tumor detection using MR images. IEEE Access, 10, 59099-59114.

Siar, M., Teshnehlab, M. (2019, October). Brain tumor detection using deep neural net-
work and machine learning algorithm. In 2019 9th international conference on computer
and knowledge engineering (ICCKE) (pp. 363-368). IEEE.

Rahman, T., Islam, M. S. (2023). MRI brain tumor detection and classification using
parallel deep convolutional neural networks. Measurement: Sensors, 26, 100694.

Rai, H. M., Chatterjee, K. (2021). 2D MRI image analysis and brain tumor detection
using deep learning CNN model LeU-Net. Multimedia Tools and Applications, 80(28),
36111-36141.
Chapter 8

Biodata

Name: Sushanth B S
Mobile No.: 8072234699
E-mail: sushanthb.sridhar@gmail.com
Permanent Address: No.64, Tivoli Justice Rathinavel Pandian Road,Golden George Na-
gar, Mogappair,Chennai- 37

Name: Harshavishal S
Mobile No.: 7338827831
E-mail: Harshavishal.s@gmail.com
Permanent Address: No.1,Vishal nagar EXTN,Saroj Nagar main road, Madipakkam,Chennai-
91

50
Chapter 8.Biodata 51

Name: Harsavardhanan K V
Mobile No.: 8072311317
E-mail: harsavardhanan@gmail.com
Permanent Address: Urbanrise Revolution One, E-Block, 17th Floor,Door No E1708,
Omr Road, Padur, Chennai-603103