A SEMINAR REPORT

on

“Multigrade Brain Tumor Classification in MRI Images”

By

Nakul Arora 122B1B010

Abhishek Chavan 122B1B047
Shivam Chavan 122B1B048
Bhavesh Chintakindi 122B1B054

Under the Guidance of

Mrs. Harshada Mhaske

TY BTech (COMPUTER ENGINEERING) 2024-25 (Semester I)

DEPARTMENT OF COMPUTER ENGINEERING,

PCET’S PIMPRI CHINCHWAD COLLEGE OF ENGINEERING
Sector No. 26, Pradhikaran, Nigdi,
Pune - 411044

PCET’S PIMPRI CHINCHWAD COLLEGE OF ENGINEERING

1
 Sector No. 26, Pradhikaran, Nigdi, Pune – 411044

DEPARTMENT OF COMPUTER ENGINEERING

CERTIFICATE

This is to certify that the seminar report entitled “Multigrade brain tumor classification in
MRI images” submitted by Name of student [PRN Number] under the supervision of Prof.
Mrs. Harshada Mhaske for the partial fulfillment of the requirement for the award of the degree of
Bachelor of Technology in Computer Engineering at Pimpri Chinchwad College of
Engineering, affiliated to Savitribai Phule Pune University, Pune in the academic year
2024-25

Prof. Mrs. Harshada Mhaske Prof. Dr. Sonali D Patil

(Seminar Guide) (Head of the

Department)

Place: Pune
Date:

2
 Acknowledgement

We express our sincere thanks to our Technical Seminar-I Guide Mrs. Harshada Mhaske for
his/her encouragement and support throughout our seminar, especially for the useful suggestions
given during the course of the seminar and having laid down the foundation for the success of this
work.
We would also like to thank our Research & Innovation coordinator Prof. Dr. Reena Kharat and
Technical Seminar Coordinator Prof. Bodireddy Mahalakshmi for their assistance, genuine support
and guidance from early stages of the seminar and during the entire course of this seminar work. We
would like to thank Prof. Dr. Sonali D Patil, Head of Computer Engineering Department for her
unwavering support during the entire course of this seminar work. We also thank all the teaching and
non teaching staff members for their help in making our seminar work successful. We also thank all
the web communities for enriching us with their immense knowledge.

Nakul Arora 122B1B010

Abhishek Chavan 122B1B047
Shivam Chavan 122B1B048
Bhavesh Chintakindi 122B1B054

3
 Abstract and Keywords

Detection of brain tumors must be accurate and timely for the purpose of proper treatment and
resultant better patient outcomes. MRI is considered the best modality for the diagnosis of brain
tumors due to its superior resolution imaging capabilities. Despite that, however, there are a number
of disadvantages regarding the perception of such MRI scans, like their manual nature and possible
variability. Thus, there exists a large interest for using deep learning models built based on
Convolutional Neural Networks in automatizing the classification of brain tumors. This still poses
challenges to existing models in terms of capturing intricate spatial relationships and high
generalization cross-dataset variations.

In this work, we propose a new hybrid model combining Swin Transformers and EfficientNetV2S to
overcome these difficulties. We use the EfficientNetV2S backbone to efficiently extract features
while Swin Transformers enhance the long-range dependencies and contextual information of brain
MRI images. We further fine-tune the proposed model using transfer learning on three different brain
tumor datasets with different classes, namely, meningioma, glioma, and pituitary tumors.

We illustrate that our hybrid approach leads to significant performance gain over state-of-the-art
approaches in terms of accuracy, precision and recall across all benchmark datasets. Moreover, the
lightweight architecture of the model ensures computational efficiency, which makes it suitable for
real-time clinical applications. This work brings a robust and scalable solution toward brain tumor
classification that will allow for further maturation in more reliable diagnostic tools of medical
imaging.

Keywords:
Detection of brain tumors, MRI, EfficientNetV2S, Swin Transformers, Deep learning, Convolutional
Neural Networks, Hybrid model, Transfer learning, Medical imaging, Tumor classification.

4
 TABLE OF CONTENTS

Sr. Contents Page. No

No.
1 Introduction
1.1 Problem Definition
1.2 Goals and Objectives
1.3 Motivation
1.4 Scope of work
1.5 Outcomes

2 Literature Survey
2.1 Summary of Literature Review
2.2 Existing System
2.3 Common findings from the Literature review

3 Algorithmic Survey
3.1 Study of Algorithms
3.2 Mathematical Model and Algorithm steps
3.3 Motivation to develop proposed system

4 Proposed System (if applicable)

4.1 Introduction
4.2 Proposed System architecture
4.3 Proposed algorithm/techniques (if any)
4.4 Proposed approach /plan of activation
4.5 (Any other information)

5 Comparative Study
5.1 Comparison Between Existing and proposed system.
5.2 Results and conclusion of Existing system

6 Advantages and Applications

6.1 Existing System Advantages / Merits
6.2 Disadvantages
6.3 Applications

7 Conclusion
7.1 Conclusion
8 References

9 Survey Paper in IEEE Format

Plagiarism report of survey paper and seminar report

5
 List of Figures

Sr. No. Figure No. Figure name Page No.

1 1.1
2 1.2
3 2.1
4 3.1
5 4.1
6 5.1
7 6.1

List of Tables

Sr. No. Table No. Table Name Page No.

1 I
2 II
3 III

6
1. Introduction
1.1. Problem Definition.

Brain tumors are malignant and non-malignant and very dangerous to human health as
these inflict severe neurological damage. Accurate detection and classification of
brain tumors are elemental for effective treatments and prognosis. Traditional
approaches use MRI scans that provide high-contrast imaging followed by either
manual or semi-automated segmentation and analysis. Recent advances in deep
learning, especially CNNs such as EfficientNet[1], have promised good state-of-the-
art results in the automation of the classification of brain tumors, yet improvement
concerning accuracy and robustness on a variety of datasets and types of tumors
seems challenging.

This current study aims to overcome these shortcomings by replacing the backbone of
EfficientNetV2S with Swin Transformers to create a hybrid model that integrates the
strengths of both architectures. EfficientNetV2S is a light efficient architecture for
feature extraction. Swin Transformers are designed to better capture long-range
dependencies and spatial relations in images. Combining those two powerful
architectures, the proposed model attempts to classify brain tumors to categories such
as meningioma, glioma, and pituitary tumors, outperforming state-of-art on various
datasets at accuracy levels. In doing so, this should provide a more accurate and
generalizable model than those already in place, hence contributing to the availability
of better diagnostic tools for clinicians and their patients.

1.2. Goals and Objectives:

Hybrid Model Design:
In this paper, we propose and design an entirely novel hybrid deep learning
architecture founded on the EfficientNetV2S and Swin Transformers. This will be
done for realizing the features extricate ability of the EfficientNetV2S from the MRI
images of the brain while using the Swin Transformers in order to capture complex
spatial relationships effectively.

High-Classification Accuracy:
The hybrid model is expected to give further distinction among meningioma, glioma,
and pituitary tumors to improve the accuracy of brain tumor classification better.
Overall performance concerning classification and pertaining to precision and recall is
better improving the classification performance on different datasets better than that
of the existing models currently have.

Optimization of model complexity and efficiency:

Achieving a balance between the complexity of models and computational efficiency
is an important goal. The objective here is to have a lean model that will deploy in
real-time clinical environments without losing performance.

Transfer Learning and Fine-Tuning:

Deploy transfer learning strategies with fine-tuning the pre-trained EfficientNetV2S
and Swin Transformers models on the target datasets related to brain tumors; this
would be a process to adapt the model towards the specific characteristics of the
medical imaging data to produce a better generalization.
7
 Generalization across Multiple Datasets:
In order to test the model's robustness and generalization, it was to be evaluated on
three independent datasets with different classifications of brain tumors. This would
have meant that the diversified imaging conditions and patient demographics are
ensured in which the model performs well.

Comparison with State-of-the-Art Models:

Benchmark the performance of the proposed hybrid model in comparison to state-of-
the-art methods in brain tumor classification, thereby showing its efficiency and
superiority.

1.3. Motivation:
The motivation for this work arises mainly from the necessity for advanced and
accurate procedures for the diagnosis and classification of brain tumors by medical
imaging. Brain tumors, benign or malignant, can cause devastating neurological
consequences as well as threaten the patient's health. Therefore, early and accurate
detection is of paramount importance in order to treat the disease properly with a good
prognosis and positive outcome of the patient. Hence, it is very important to develop
strong diagnostic tools which can guide clinicians to make appropriate decisions.

Recent breakthroughs in deep learning with the use of CNNs have hinted at very
exciting prospects in the automatic interpretation of medical images. Even though
huge steps have been taken forward, the research still faces some challenges in
achieving greater-than-average classification accuracy and robustness across various
datasets with distinct types of tumors. Because the existing models fail to capture
intricate spatial interactions present in MRI images, it has led to high misclassification
rates, affecting healthcare at the end.

This work is motivated by the need to overcome the following limitations by

combining two state-of-art architectures: feature extraction efficient, capable
architectures, such as EfficientNetV2S, and dependency-capturing and
contextualizing architectures, like Swin Transformers. The proposed approach
combines both architectures into a hybrid model with improved classification and
generalization abilities while maintaining computational efficiency.

Such advanced methods would have a beneficial impact in being included in clinical
workflows, which would leave a relatively light burden on radiologists and increase
the possibility of well-timed treatments. This is why our research interest is prompted
by the desire to contribute to the large number of medical imaging and deep learning
bodies of knowledge, opening avenues for further research and innovation in this
crucial area.

It is this move to enhance the patient outcome in better diagnosis accuracy, the
promotion of innovative approaches in brain tumor classification, and commitment to
bridging the gulf between advanced technology and clinical application in health that
motivates this research.

8
1.4. Scope of work:
Review and Dataset Analysis:
This is based on conducting an extensive review of the current state-of-art deep
learning models applied to brain tumor classification tasks. Analyze various available
brain tumor MRI datasets regarding size and class distribution and associated imaging
modalities to understand various challenges and requirements to develop models.

Data Preprocessing and Augmentation:

While training, data-preprocessing techniques can be used on MRI images like
normalization, resizing, and augmentation. It handles the imbalanced data set by
removing artifacts and generally augments the dataset. Usually, it does that with
procedures like rotation, flipping, and scaling to make the model robust.

Model Design and Integration:

Hybrid Architecture Designed: EfficientNetV2S + Swin Transformers Experiment
several variations and integration strategies, like using Swin Transformer as an
attention module or feature extractor, to find the best design.

Transfer Learning and Fine-Tuning:

Use pre-trained EfficientNetV2S and Swin Transformers models followed by fine-
tuning on the brain tumor datasets. The training process might be improved by using
some techniques, such as layer freezing, learning rate scheduling, regularization.
The hybrid model will be trained on the prepared datasets using the right loss
functions and the metrics of evaluation. The model will be validated by employing
cross-validation techniques, where it can check if the model is reliable and generalizes
well to unseen data.

Performance Evaluation and Benchmarking:

To measure the performance of the model, apply error metrics like accuracy,
precision, recall, F1-score, and AUC-ROC and compare it with the baselines. Use
isolated efficientnet V2S, isolated Swin Transformers or other state-of-the-art
methods which are proper to validate the proposed hybrid approach.

Hyperparameter Optimization:
Much hyperparameter tuning should be done to make the model better. This includes
learning rate, batch size, optimizer setting, and depth of network.

Interpretability and Visualization:

Through the application of techniques, such as Grad-CAM or SHAP, the brain MRI
images can be visualized, and one can pinpoint where most of the model has been
deciding on. That will make interpreting what the model predicts feasible and reliable
enough for adopting it into the clinical setting.

Real-World Deployment and Testing:

Real-World Clinical Setting Feasibility Studies: This should show the applicability of
the model to the clinical setting. Test the system outside the development
environment on test data or real-time MRI scans for the determination of the practical
feasibility of the software.

9
 Documentation and Reporting:
During the research process, one has to document methodology, experiments, results,
findings, and everything else relating to a research. This, along with the
comprehensive report, may be published in a peer-reviewed journal or presented in
conferences relevant to the subject.

1.5. Outcomes:
With EfficientNetV2S and Swin Transformers combined, it would be possible in the
proposed research work to develop a hybrid deep learning model that enhances brain
tumor classification. The outcome of the work resulted in the following:

Enhancement of Classification Accuracy:

Accuracy, precision, and recall obtained in three-class classification of the brain
tumors by classifying them into one of three categories, namely, meningioma, glioma,
or pituitary tumors, have reached higher accuracies.
A comparison of the hybrid method on each of the test sets with the state-of-the-art
techniques clearly shows that the hybrid method has been clearly more efficient on all
the datasets, which obviously depicts its suitability for clear differentiation among
various types of tumors.

Improved Generalization:
As the model is capable of generalization over the different datasets, including
different imaging conditions and patient demographics, integrating Swin
Transformers with EfficientNetV2S allows the model to seize local and global
features from MRI images.

Computational Efficiency:
The hybrid model still maintains a light architecture despite improved performance.
The model, therefore, remains practicable for deployment in clinical environments in
real time. This efficiency ensures that the model would be easily integrated into the
established medical imaging workflows with minimal computational overhead.

Clinical Use:
It is quite accurate and efficient, thereby standing as a good tool that may help
radiologists in making early detection and classification of brain tumors in order to
reduce the time usually taken to diagnose and thereby plan better for treatment.

Scalable Framework:
The proposed study provides a hybrid approach as a scalable framework, which can
be extended to other medical imaging tasks apart from the classification of brain
tumors. It therefore opens up new ways for the integration of CNNs and Transformers
in medical applications.

2. Literature Survey:

A great deal of research has been conducted recently in the field of brain tumor classification using
deep learning models, especially after the introduction of advanced architectures like DenseNet,
EfficientNet, and VGG.
10
One outstanding way is the application of DenseNet-161 to brain tumor classification, which gave a
correctness of 99.25% on MRI images from 233 patients. Nevertheless, it was a method that was
fine-tuned over 25 epochs but it had some difficulties such as modified shape of tumor and
overlapping intensity values causing the quality of the image to be affected.

Another study used EfficientNetV2S that was carried out through the transfer of learning of images.
The problem had accuracies from 95.6% to 98.48% in different datasets. Yet, this method was
tangled up in the fact that only MRI images with labels were available making the situation complex.

On a different note, a hybrid model was built which is composed of ResNet50 and ConvNeXt to get
passed this club of manual brain tumor classification. To be precise, it combined the feature
extraction of ResNet and the faster computation of ConvNeXt so that it can get the accuracy and
robustness of brain tumor identification.

One of the prior methods such as CNN-SVM combinations brought 95.82% of accuracy when
organizing small datasets and overfitting was decreased. Methods based on transfer learning also
worked well like EfficientNetV2S, and using deep learning algorithms, as VGG-16 was able to
classify tumors up to 91.4% accuracy despite the issues with unbalanced data.

Another interesting model is the VGG16-NADE hybrid. This model joins VGG-16 to feature
extraction and NADE to noise abatement and density estimation technology to achieve 96.01% of
precision. There were some issues that the model encountered such as for instance, the dataset was
not uniform so that it differed from each other, and also as the use of additional deep convolutional
layers increased the computational cost. Moreover, learning methods with VGG-16 that were applied
on a smaller dataset showed a reduction of 42 points off accuracy having a 52.47% accuracy, while
on a bigger dataset multi-class classification using VGG-16 managed to reach the highest 98.62%
efficacy.

Even though their efficiency has become better, some troubles still exist like overfitting, not enough
labeling data, and unsatisfactory generalization of diverse types of tumors.

3. Algorithmic Survey:

4. Proposed System:

5. Comparative Study:

6. Advantages and Applications:

Some of the Advantages of our Proposed System are as follows:

○ High Accuracy: The EfficientNet model offers very accurate classification results,
ensuring reliable diagnosis of different types of brain tumors.
○ Early Detection: Its ability to classify tumors quickly allows for early diagnosis, which is
11
 critical for effective treatment.
○ Efficiency: Automating the classification process makes it faster, reducing the time
required for doctors to get diagnostic results.
○ Consistency: Provides consistent and repeatable results, eliminating errors caused by
manual interpretation.
○ Explainability: Grad-CAM visualization helps radiologists understand the model’s
decisions by highlighting areas of the MRI scan that are most indicative of a tumor. ○
Scalability: The model can handle large datasets, making it ideal for busy hospitals or
diagnostic centers with large numbers of MRI scans.

The Applications of our Proposed System are as follows:

○ Hospitals and Diagnostic Centers: Assists doctors in diagnosing and grading brain
tumors more accurately and quickly.
○ Treatment Planning: Helps in making informed treatment decisions based on accurate and
detailed tumor classification.
○ Medical Software: Can be integrated into existing MRI analysis systems for real-time
tumor classification.
○ Research: Supports research by providing a platform for analyzing large datasets of
brain MRI scans to study tumors.
○ Remote Healthcare: Enables accurate diagnosis in telemedicine, where medical services
are provided remotely, allowing doctors to assist patients in different locations.

7. Conclusion:

The research paper "Multigrade Brain Tumor Classification in MRI Images" mentioned a hybrid
model made up of EfficientNetV2S and Swin Transformers which are used for the development of
brain tumor classification in MRI scans. EfficientNetV2S is the protocol that efficiently helps in the
extraction of features while Swin Transformers are the architecture that captures spatial relationships
across layers, resulting in improved classification accuracy, precision as well as recall of tumors such
as meningioma, glioma, and pituitary tumors. Particularly, the model is light-weight, thus, it is
suitable for real-time clinical applications and demonstrates strong generalization across diverse
datasets. Once its features are properly inferred it can be re-scaled for multiple medical imaging jobs.
To sum up, this model that was suggested has to be the best, most efficient solution that not only
outperforms existing methods but also is a real tool for improving diagnostic accuracy in healthcare.

8. References:

[1] Multigrade Brain Tumor Classification in MRI images using Fined-Tuned Efficient Net
(Link)

12
[2] DenseNet for Brain Tumor Classification in MRI Images (Link)

[3] Brain Tumor Classification Using Convolutional Neural Network (Link)

[4] Grade Classification of Tumors from Brain Magnetic Resonance Images Using a Deep
Learning Technique (Link)

[5] Automated Brain Tumor Detection From Magnetic Resonance Images Using Fine-
Tuned EfficientNet-B4 Convolutional Neural Network (Link)

[6] Brain Tumor Classification using Deep Learning Framework (Link)

[7] Classification using deep learning neural networks for brain tumors (Link)

[8] Brain Tumor Identification and Classification of MRI images using deep learning
techniques (Link)

[9] A Novel Deep Learning Method for Recognition and Classification of Brain Tumors
from MRI Images (Link)

13