K.L.N.

COLLEGE OF ENGINEERING
An Autonomous Institution, Affiliated to Anna University Chennai

DEPARTMENT OF ELECTRONICS AND COMMUNICATION ENGINEERING

Brain Tumor Detection using Efficient Net B3

and YoloV8
Domain: Digital Image Processing

Supervisor Team Members

Mrs. R. Angayarkanni G. Sankarram (910621103073)
Assistant professor/ECE P.G. Sharann (910621103078)
S. Surya Prakash(910621103088)
 INTRODUCTION

• It focuses on detecting and classifying brain tumors in MRI images using the
YOLOV8 (You Only Look Once)
• Develop an efficient, real-time detection system for identifying different types of
brain tumors from MRI scans, aiding in early diagnosis
• However, manual interpretation of MRI scans is time-consuming and prone to
human error.
• This study explores the application of the YOLOv8 model, leveraging its advanced
object detection capabilities for accurate and rapid brain tumor detection and
classification in MRI images.
 OBJECTIVE
• Detection -- Accurately identify brain tumors in MRI images and localize them with

bounding boxes.

• Classify the detected tumor -- into relevant categories (glioma, meningioma (small,

medium, large), pituitary tumor , No tumor)

• Optimize the model -- for real-time processing, ensuring its applicability in clinical

settings where quick diagnosis is crucial.

• Validation and Evaluation – To assess the performance of the YOLOv8 model in

comparison to existing techniques by measuring key metrics such as accuracy,

precision, and recall. These evaluation results are accessible through the web interface

to demonstrate model effectiveness.

 ABSTRACT
• Brain tumor detection is a critical task in medical image analysis, where early and
accurate identification can significantly improve patient outcomes.
• EfficientNet-B3 with the object detection capabilities of YOLOv8 for precise
localization and classification of brain tumors from MRI scans.
• EfficientNet-B3 is employed to extract high-level semantic features due to its
optimized depth, width, and resolution scaling.
• YOLOv8 leverages these features for real-time tumor detection and bounding box
prediction.
• The proposed model was trained and evaluated on a curated dataset of brain MRI
images, achieving high accuracy, precision.

4
 LITERATURE SURVEY
S.N Title of the Author Journal / Year of Summary Benefits Drawbac
o Paper Conferenc Publicati of the ks
e Name ons paper

1 Detection and M.S. Journal of December YOLO The Ensuring

classification on Mithun Radiation 2024 NAS YOLO the model
MRI images of and S. Research model NAS deep is trained
brain tumor Joseph and achieved learning on a
using YOLO Jawhar Applied high model, diverse
NAS deep Sciences accuracy which is dataset to
learning mode in both known for generalize
detection its real- well
and time across
classificat object different
ion tasks. detection patient
capabilitie demograp
s, to hics and
achieve MRI
this goal. machines

5
 LITERATURE SURVEY
S.No Title of the Author Journal / Year of Summary Benefits Drawbacks
Paper Conference Publicat of the
Name ions paper

1 Deep learning- Ari, A., Turkish 2018 It presents It showed Many older
based brain & Journal of a novel how studies
tumor Hanbay, Electrical approach effective (likely
classification D. Engineering to assist deep including
and detection and radiologist learning this one)
system” Computer s in (especially used
Sciences, detecting CNNs) can relatively
and be in small
classifyin medical datasets,
g brain diagnosis, which can
tumors specifically affect
using for brain generalizati
cranial tumor on to real-
MRI detection. world
images. scenarios

6
 LITERATURE SURVEY
S.No Title of the Author Journal / Year of Summary Benefits Drawbac
Paper Conferen Publicatio of the ks
ce Name ns paper

2 Brain MR Latif, G., Current 2021 The study The The

Image Iskandar, Medical concludes proposed method
Classification D. A., Imaging that the method only
for Glioma Alghazo, (Volume proposed achieved classifies
tumor J., & Butt, 17, Issue method 98.77% images
detection using M. M. 1, Pages offers accuracy, and does
deep 56–63). superior surpassing not detect
convolutional accuracy traditional or locate
neural network in brain approache the tumor
features MRI s in within the
classificati glioma image (no
on and detection. bounding
glioma boxes or
tumor segmentati
detection on).
 Proposed Methodology
• Data collection :MRI Brain scan images are gathered from Kaggle. The dataset includes
different tumor types such as glioma, meningioma, and pituitary tumors.

• Preprocessing :The dataset undergoes preprocessing including resizing, normalization and

filtering to enhance image quality.

• Feature Extraction and Annotation : EfficientNet-B3 is employed for feature extraction

from the preprocessed images. After feature extraction, a heat map is generated using Grad-
CAM, which highlights the tumor regions by annotation and labeling of the highlighted
regions

• Model Training: The YOLOv8 model is trained using PyTorch for optimized performance.
An 80% training and 20% testing split is applied to dataset for accurate detection.

• Model Evaluation: The trained YOLOv8 model is evaluated using metrics like mAP, IoU,
Precision, and Recall to assess detection accuracy.

• Inference & Deployment: YOLOv8 runs inference on new MRI images can be run in a
Jupyter Notebook for development and testing. In addition, web interface is built for
 Differences between YOLO-NAS and
YOLOv8
Aspect YOLO NAS YOLOv8

Full Form YOLO Neural Architecture You Only Look Once version
Search 8

Architecture designed using Neural designed by the Ultralytics

Architecture Search (NAS) team

Accuracy Slightly higher in some tasks Very high accuracy, state-of-

due to customized NAS- the-art in most use cases
generated layers

Speed Fast, but generally slower Very fast, suitable for real-
than YOLOv time detection

Efficiency in Use Efficient, depending on NAS- Extremely efficient and

generated structure lightweight for deployment

04/21/2025 9
 Outcome Of The Project

• The model will accurately classify brain tumors into specific types, such as
gliomas, meningiomas, and pituitary tumors.
• The proposed brain tumour detection and classification system leverages the
YOLOv8 model to provide an efficient, real-time solution for analysing MRI
images
• It accessed through a web interface, with results can be viewed in real time.
• Efficient Net B3 extracts deep, multi-scale features from MRI images, making
detection more reliable even in complex cases.
PROPOSED METHODOLOGY:BLOCK
DIAGRAM
 Explanation of Dataset
• MRI brain tumour Dataset obtained from Kaggle

classification No of images

Total images 2690

Training ( 80%) 2152

Testing (20%) 538

Glioma Tumour 746

Meningioma Tumour 810

Pituitary Tumour 807

No Tumour 327
Training of yolov8 model
 Explanation of unique concepts used in
the project
• Internal Classification of Meningioma Tumor: Meningioma tumor are classified
into small, medium, large based on their size

• Visual Interpretability using Grad-CAM: Grad-CAM (Gradient-weighted Class

Activation Mapping) is used to generate heatmaps that highlight the regions of the
image the model focuses on while detecting tumors which helps tio validate model
predictions by localizing tumor regions

• Web-based Interface: Users can easily access and utilize the brain tumor detection
model through a convenient web interface

04/21/2025 14
Evaluation of yolo v8
 Evaluation of yolov8

Metric / Loss Description Formula / Explanation

train/box_loss Bounding box regression loss during training.

train/cls_loss Classification loss during training.

train/dfl_loss Distribution Focal Loss for bounding box

prediction.
val/box_loss Bounding box regression loss on validation data. Same as train/box_loss, but on validation
set.
val/cls_loss Classification loss on validation data. Same as train/cls_loss, but on validation
set.
val/dfl_loss Distribution Focal Loss on validation data. Same as train/dfl_loss, but on validation
set
metrics/ Precision: predicted positives that are true
precision(B) positives (class B).
metrics/recall(B) Recall: actual positives that are correctly
predicted (class B).
metrics/mAP50- Mean Average Precision over IoU thresholds from
95(B) 0.50 to 0.95 (class B)

04/21/2025 16
 Software Requirements
• Python 3.9
• Jupyter Notebook
• HTML,CSS - frontend
• Flask/Fast API – backend

04/21/2025 17
 Experimental Results
Category Original image Output image by the trained
model
glioma

Meningioma-small

Meningioma-medium

Meningioma-large

Pituitary

No tumour

04/21/2025 18
 Experimental results

04/21/2025 19