Project Report

On
Brain Tumer Detection Using CNN

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING

SIKKIM MANIPAL INSTITUTE OF TECHNOLOGY
(A constituent college of Sikkim Manipal University)
MAJITAR, RANGPO, EAST SIKKIM- 737136

In partial fulfilment of requirements for the award of degree in

Bachelor of Technology in Computer Science and Engineering (Artificial Intelligence
and Machine Learning)
(2024)

By
Pawan Dahal(202200540)
Devashish Khawas(202200099)
Deepan Sharma(202200475)
Niket Kumar Gupta(202200525)
Tejas Kumar(202200486)

Under the supervision of:

Dr. Basab Nath
Assistant Professor, Dept of CSE, SMIT

SL. Content Page No.

1|Page
NO.

1 Abstract 3

2 Introduction 4-5

3 Literature Survey 6-9

4 Problem Definition 10-11

5 Proposed Solution Strategy 12-13

6 Methodology 14-16

7 Flowchart 17

8 Gantt Chart 18

9 Conclusion 19

10 References 20

Table of Contents

2|Page
 Abstract
 This project focuses on the development of a deep learning model

using Convolutional Neural Networks (CNNs) for the classification

and segmentation of brain tumors in MRI images.
 The objective is to first classify MRI scans as either tumor-
positive or tumor-negative, and for positive cases, segment the
tumor region within the image.
 By combining classification and segmentation tasks in a multitask
learning framework, the model offers a more efficient and
comprehensive solution for both quick detection and precise
localization of brain tumors.
 This dual approach can assist healthcare professionals in early
diagnosis and treatment planning, enhancing the overall accuracy
and speed of tumor detection..

3|Page
 Introduction

Brain tumors are among the most critical medical issues, requiring
accurate and early detection to improve patient outcomes. MRI (Magnetic
Resonance Imaging) is the go-to tool for identifying brain abnormalities,
thanks to its ability to provide detailed views of brain structures.
However, manually reviewing MRI scans can be slow, subjective, and
prone to human error, which can lead to delayed diagnoses and
inconsistent results.

In this project, we’re focusing on brain tumors that fall into three key
categories: benign (non-cancerous but can still cause harm by pressing on
surrounding tissues), pre-malignant (which could eventually become
cancerous), and malignant (cancerous and often aggressive, requiring
immediate attention). Accurately identifying and localizing these tumors
is crucial for deciding the right treatment, whether it's surgery, radiation,
or other therapies.

With advances in artificial intelligence, particularly deep learning, we're

able to improve the speed and accuracy of tumor detection. Convolutional
Neural Networks (CNNs), designed to analyze images, are a powerful
tool for automating this process. In this project, we’re building a CNN
model that combines two important tasks: classification (to detect
whether or not a tumor is present) and segmentation (to outline the exact
boundaries of the tumor). This multitask approach allows us to not only
identify the presence of a tumor but also provide detailed information
about its location, helping healthcare professionals make faster, more
precise decisions.

4|Page
By integrating classification and segmentation in one model, we aim to
streamline the diagnostic process, saving time while improving accuracy.
This could make a significant difference in treatment planning, especially
for malignant tumors that require immediate attention. Moreover,
automated detection reduces variability in diagnoses across different
medical experts, ensuring that patients receive more consistent care.

Ultimately, this project aims to bring the power of CNNs into clinical
practice, offering doctors an efficient and reliable tool to enhance brain
tumor detection and improve patient care. By focusing on both the
presence and location of tumors, we’re giving medical professionals the
detailed, actionable insights they need to provide better, faster treatment.

5|Page
 Literature Survey

AUTHOR NAME, TITLE OF FINDINGS RELEVANCE

JOURNAL NAME, PAPER TO THE
VOL., PAGE NO., PROJECT
YEAR

1. Arkapravo MRI-based The paper looks at The paper

Chattopadhyay,Mausum brain tumour different CNN provides
i Maitra. Neuroscience image models for detecting effective
Informatics(2022) detection using brain tumors in MRI techniques for
CNN based scans, showing how classification
deeplearning they outperform and,segmentation
method.[1] traditional methods tasks. It
by automating feature emphasizes
extraction and automation of
boosting accuracy. tumor detection
It also highlights and enhancement
how these models of diagnostic
help doctors with accuracy,
early diagnosis and aligning with our
treatment planning by objectives. The
reducing human challenges and
error. methodologies
However, challenges discussed will
like tumor variability, guide our design
differences in choices, helping
imaging, and the high us implement
demands of real-time suitable
detection still make algorithms for
large-scale use efficient
difficult, even though detection and
the results are localization in
promising. MRI images.

AUTHOR TITLE OF FINDINGS RELEVANCE

NAME, PAPER TO THE

6|Page
 JOURNAL PROJECT
NAME, VOL.,
PAGE NO.,
YEAR

2. Mesut Toğaçar, BrainMRNet: The paper introduces a The relevance of

Burhan Ergen, Brain tumor CNN model for brain this paper to our
Zafer Cömert, detection using tumor detection in MRI project is that it
Medical magnetic scans, showing improved details
Hypotheses resonance accuracy over traditional significant
Volume: 132 images with a methods. It emphasizes the improvements in
109031, 2020 novel model's role in early
recognition
convolutional diagnosis and treatment
accuracy for
neural network planning but
model[4] acknowledges challenges brain tumor
. like tumor variability and detection. It also
high computational discusses major
demands. challenges faced
during
implementation,
providing
insights that will
inform our
approach.
Additionally, it
emphasizes the
importance of
training with a
properly sized
dataset, which is
crucial for
achieving
reliable results
in our
classification
and
segmentation
tasks.

7|Page
 AUTHOR TITLE OF FINDINGS RELEVANCE
NAME, PAPER TO THE
JOURNAL PROJECT
NAME, VOL.,
PAGE NO.,
YEAR

3. Hossain, M.S., Automated The paper presents The relevance of

Mollah, M.N. Brain Tumor methods for automated this paper to our
IEEE Access, Detection and brain tumor detection and project lies in its
2021. Segmentation segmentation using deep detailed
Using Deep learning techniques. It exploration of
Learning“ demonstrates that deep deep learning
learning models, techniques for
particularly CNNs, tumor detection
significantly enhance and
detection accuracy and segmentation. It
speed compared to highlights
traditional methods. The improvements in
study also addresses accuracy that
challenges like data align with our
variability and goals.
emphasizes the Additionally, the
importance of using emphasis on
diverse datasets for using diverse
effective training. datasets will
guide our
training strategy,
ensuring robust
performance in
detecting and
localizing brain
tumors in MRI
images.

8|Page
 AUTHOR TITLE OF FINDINGS RELEVANCE TO
NAME, PAPER THE PROJECT
JOURNAL
NAME, VOL.,
PAGE NO.,
YEAR

4. Md. Saikat Islam Accurate brain The study shows The paper is relevant to
Khan, Anichur tumor detection high accuracy in our project as it
Rahman, Tanoy using deep detecting and highlights the use of
Debnath, Md. convolutional classifying brain CNNs for improving
Razaul Karim, neural network tumors using MRI detection accuracy. Its
Mostofa Kamal data and deep methodologies on data
Nasir, learning preprocessing and
Computational techniques. A 23- segmentation techniques
and Structural layer CNN was provide valuable
Biotechnology developed, with insights to enhance my
vol 20, transfer learning system's design and
2022 used to address effectiveness in tumor
overfitting on detection and
smaller datasets. segmentation.
The findings
indicate improved
diagnostic
precision and
efficiency, with
potential real-time
medical
applications. The
models and
datasets are
publicly available
for further
research.

9|Page
 Problem Definition

Brain tumors present significant challenges in medical imaging due to

their varying sizes, shapes, and positions within the brain. Accurate
detection and localization of tumors in MRI scans are critical for
diagnosis, treatment planning, and surgery. Traditional methods rely
heavily on manual analysis by radiologists, which is labor-intensive and
subject to human error. Moreover, single-task CNNs typically focus on
either classification (detecting the presence of a tumor) or segmentation
(localizing the tumor), making them insufficient for real-time medical
decision-making.

In this project, we are addressing the complexity of automatic

segmentation, particularly for gliomas—a type of tumor that occurs in the
brain and spinal cord. Gliomas are especially challenging to segment due
to the nature of MRI data, which can be complicated by differences in
devices and acquisition protocols. These variations create intensity biases
and other inconsistencies across datasets, making it harder for models to
accurately segment tumors. Moreover, the process is further complicated
by the need to segment different tumor sub-regions, each of which may
require distinct modalities to achieve accurate results.

Key problems we aim to solve include:

10 | P a g e
1. Classification alone is insufficient: While detecting the presence of a
tumor is important, it doesn’t offer enough information for effective
treatment planning. Medical professionals need to understand the size,
shape, and location of the tumor to make informed decisions about
surgery or therapy.

2. Segmentation alone is computationally intensive: Segmentation

models, while precise, require significant computational resources and
time to process each image, limiting their practicality in real-time clinical
settings where quick decisions are crucial.

3. Need for efficient multitask learning: To overcome these challenges,

we propose a multitask CNN that can handle both classification and
segmentation simultaneously. This approach optimizes both speed and
accuracy, providing a comprehensive solution for real-time medical
decision-making. By combining these tasks, the model can quickly detect
tumors while also delivering precise information about their location and
characteristics, reducing the burden on radiologists and improving patient
outcomes.

By tackling the complexities of glioma segmentation and developing an

efficient multitask framework, this project aims to streamline the
diagnostic process, enhancing the speed and accuracy of brain tumor
detection and treatment planning

11 | P a g e
 Problem solving strategies

In this project, we propose a dual-task learning approach using a CNN

architecture capable of performing both tumor classification and
segmentation simultaneously. By utilizing a multitask loss function that
combines binary cross-entropy for classification and dice loss for
segmentation, we aim to optimize both tasks in a single training process.
The model is designed to be efficient for real-time use, providing critical
tumor detection and localization, which will assist radiologists and
surgeons in treatment planning. This approach addresses the limitations
of traditional methods that focus on either classification or segmentation,
offering a more comprehensive solution for brain tumor analysis.

Key Problems:

1. Classification alone is insufficient: Classification models focus

solely on detecting the presence of a tumor without providing
necessary spatial details, such as size, shape, or location. While
tumor detection is important, it is not enough for comprehensive
treatment planning, where precise localization is critical. This
limitation can hinder surgical or therapeutic decisions, making
classification alone inadequate for medical decision-making.
Therefore, combining classification with segmentation is essential
for providing more actionable insights for treatment.

2. Segmentation alone is computationally intensive: While

segmentation models offer precise localization of tumors, they are
computationally demanding, as each pixel in an image must be
analyzed to determine whether it belongs to the tumor. This high
computational cost makes segmentation models less suitable for

12 | P a g e
 real-time applications, particularly in clinical settings where fast
decision-making is essential. The need for significant processing
time and resources limits their scalability, making segmentation
alone impractical for high-volume workflows in hospitals and
medical centers.

3. Need for efficient multitask learning: Using separate models for

classification and segmentation creates inefficiencies, as
radiologists must rely on multiple systems or manually integrate
outputs to make comprehensive decisions. Multitask learning
addresses this by integrating both tasks into a single CNN
architecture, which optimizes the speed and accuracy of both tumor
detection and localization. This combined approach eliminates the
need for manual intervention and improves workflow efficiency,
making it ideal for real-time use in medical environments.

13 | P a g e
 METHODOLOGY

1. Data Collection & Preprocessing:

 Dataset Acquisition: We will collect publicly available brain MRI

datasets from platforms such as Kaggle, ensuring they include
labeled images suitable for both tumor classification and
segmentation tasks.

 Image Preprocessing:

o Resizing: All images will be resized to a uniform dimension

(e.g., 256x256 pixels) to standardize input for the CNN
model.

o Normalization: Pixel values will be normalized to a range of

[0, 1] to enhance model convergence.

o Data Augmentation: To increase dataset diversity and reduce

overfitting, we will apply data augmentation techniques such
as:

 Random rotations (e.g., 90°, 180°, 270°)

 Horizontal and vertical flips

 Zoom adjustments

 Brightness and contrast changes

2. CNN Architecture:

14 | P a g e
  Model Design:

o Shared Feature Extraction: A shared convolutional network

will extract relevant features from MRI images, supporting
both classification and segmentation tasks.

o Classification Head: One branch of the network will classify

whether a tumor is present, using fully connected layers to
output binary predictions.

o Segmentation Head: Another branch will handle

segmentation, outputting a pixel-wise mask to localize tumor
regions. This will involve upsampling techniques and skip
connections to maintain spatial accuracy.

Training the Model:

 Multitask Learning: The model will be trained to optimize both

classification and segmentation tasks simultaneously. A multitask
loss function combining binary cross-entropy (for classification)
and dice coefficient loss (for segmentation) will guide training.

 Data Splitting: The dataset will be split into training (80%) and
validation (20%) subsets. Validation data will help assess model
performance during training.

 Training Process: An optimizer like Adam will be used with an

appropriate learning rate schedule. Early stopping will monitor
validation loss to prevent overfitting.

Evaluation:

 Performance Metrics:

15 | P a g e
 o Classification: Metrics such as Accuracy, Precision, Recall,
and F1 Score will evaluate the model’s ability to detect
tumors.

o Segmentation: Metrics like Dice coefficient and Intersection

over Union (IoU) will assess how well the model segments
tumor regions.

 Visualization: Predicted segmentation masks will be visually

compared to ground truth masks to confirm qualitative accuracy.

Real-world Application:

 Deployment: The final trained model will be deployed as a clinical

decision support tool for radiologists, allowing for automatic
detection and localization of tumors from MRI scans.

 Feedback and Improvement: Radiologists will be trained on how to

use the system, and feedback will be gathered to refine the model
for enhanced real-world performance.

16 | P a g e
FLOWCHART

17 | P a g e
GANTT CHART

18 | P a g e
 CONCLUSION
This project outlines the potential of Convolutional Neural Networks
(CNNs) in enhancing brain tumor detection in MRI images. While the
model is yet to be developed, our proposed approach integrates both
classification and segmentation tasks into a single framework, aiming to
improve diagnostic accuracy and streamline the process for healthcare
professionals. By reducing the need for manual analysis and providing
real-time tumor localization, this system has the potential to assist
radiologists and surgeons in making more informed treatment decisions.

19 | P a g e
 References

[1].A. Chattopadhyay and M. Maitra, "MRI-based brain tumour image

detection using CNN based deep learning method," Neuroscience
Informatics, vol. 2, p. 100060, 2022.
[2]D. Lamrani, B. Cherradi, O. El Gannour, M. A. Bouqentar, and L.
Bahatti, "Brain Tumor Detection using MRI Images and Convolutional
Neural Network," International Journal of Advanced Computer Science
and Applications (IJACSA), vol. 13, no. 7, pp. 452-457, 2022, doi:
10.14569/IJACSA.2022.0130755.
[3]M. S. I. Khan, A. Rahman, T. Debnath, M. R. Karim, M. K. Nasir, S.
S. Band, A. Mosavi, and I. Dehzangi, "Accurate brain tumor detection
using deep convolutional neural network," Future Technology Research
Center, National Yunlin University of Science and Technology, vol. 2022,
pp. 1-8, Aug. 2022.
[4]M. Toğaçar, B. Ergen, and Z. Cömert, “BrainMRNet: Brain tumor
detection using magnetic resonance images with a novel convolutional
neural network model,” Med. Hypotheses, vol. 134, p. 109531, Jan. 2020.

20 | P a g e