Detection of Brain Tumors Using Advanced Image Processing and

the Ensemble Model and YOLO Family

P. Poorna Priya and N. Vidhya Sree

Department of Electronics and Communication Engineering, DIET, Anakapalle, Visakhapatnam, Andhra Pradesh, India

Keywords: You Only Look once, CNN, Image Processing, Brain Tumor, Machine Learning, Deep Learning, Tumor

Detection, Classification.

Abstract: Effective diagnosis and therapy of brain tumors depend on their identification and classification. Using the

Brain Tumor dataset, this study makes use of sophisticated transfer learning models and deep convolutional

neural networks (DCNNs). When models like DCNN, ResNet152, EfficientNetB2, Exception, and

Nonmobile were tested, an ensemble of Exception and Nonmobile produced the best accuracy (98.1%). Grade

0 (no malignancy) to Grade III (big tumor) were the four grades into which tumors were divided. With a mean

average precision (map) of 78.9%, YOLOv9 fared better than other models for anomaly detection. A Flask-

based interactive interface was created for safe and easy access in order to improve usage.

1 INTRODUCTION

Unusual cell growths in the brain called brain tumors

raise the pressure inside the skull, impairing essential

processes including movement, speech, and thought.

They are categorized as either malignant (aggressive,

fast-growing, and invasive) or benign (slow-growing,

less invasive), and early identification is essential for

successful treatment. Conventional diagnosis uses

MRI and CT scans, which need to be interpreted by

experts and can be laborious and error-prone.

Improvements in deep learning (DL), especially

convolutional neural networks (CNNs), and artificial

intelligence (AI) have greatly enhanced tumor

categorization and detection (Kumar et al., 2022;

Ullah et al., 2022; Babu Vimala et al., 2023). Better

patient outcomes result from these models increased

diagnostic precision, accelerated analysis, and

support for early action. In order to increase

precision, effectiveness, and clinical decision-

making, this study investigates the use of deep

learning and machine learning for automated brain

tumor identification (Mathivanan et al., 2024; Das &

Goswami, 2024).

2 RELATED WORKS

MRI scans have been used in a number of researches

to investigate AI and deep learning methods for brain

tumor identification. Asif et al. used pre-trained CNN

architectures such as VGG16, Reset, and Inception to

introduce transfer learning-based models and show

increased classification accuracy. By improving CNN

models and focusing on feature extraction for

improved tumor classification, Srinivas et al. further

improved this methodology. CNN-based transfer

learning was used by Bairagi et al., who demonstrated

how effective it is at processing intricate medical

images. In order to improve detection accuracy with

little data, Anjum et al. optimized pre-trained

networks using ResNet50 and InceptionV3. By

comparing several CNN models, Khaliki and

Başarslan were able to verify that transfer learning

performed better in terms of classification accuracy

and computational efficiency than traditional three-

layer CNNs. By combining DenseNet169 with

machine learning classifiers such as Random Forest

and SVM, Khan et al. developed Crossover NET,

which improved tumor classification. CNNs and

transfer learning were integrated by Incir and

Bozkurt, who showed enhanced performance on

sizable and varied datasets. In order to highlight the

significance of flexible AI models in medical

imaging, Sadad et al. expanded deep learning

628

Priya, P. P. and Sree, N. V.

Detection of Brain Tumors Using Advanced Image Processing and the Ensemble Model and YOLO Family.

DOI: 10.5220/0013939700004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 5, pages

628-633

ISBN: 978-989-758-777-1

applications to multi-class tumor classification

(benign, malignant, and metastatic) using VGGNet

and ResNet.

3 MATERIALS AND METHODS

The proposed brain tumour detection and

characterization framework fosters areas of strength

for a for-brain tumour identification and grouping by

utilization of a Brain Tumour dataset. To further

develop grouping accuracy, the framework totals

“deep convolutional neural networks (DCNNs)” with

cutting edge move learning models. It uses a half

breed model of DCNN and ResNet152 serving as the

baseline for growth grouping, utilizing modern

models like "DCNN, ResNet152, EfficientNetB2,

Exception, and Nonmobile." The four forms of cancer

are classified as "Grades 0 (no tumor), Grade I (little

tumor), Grade II (medium-sized tumor), and Grade III

(big tumor)”. To detect abnormalities in brain scans,

the invention makes use of cutting-edge YOLO

models, such as "YOLOv5x6, YOLOv5s6, YOLOv8,

YOLOv9."These models are trained and optimized to

enable precise cancer identification and

classification. Incorporated utilizing a Flask based

intelligent point of interaction, this offers an easy to

understand stage for clinical applications and safe

client recognizable proof.

Figure 1 shows The

Proposed Architecture.

Figure 1: The proposed architecture.

By integrating data augmentation, image

processing techniques, and deep convolutional neural

networks (DCNNs) for object detection and picture

sequencing, this architecture (fig.1) develops a deep

learning-based framework for image analysis.

Utilizing YOLO models “YOLOv5x6, YOLOv5s6,

YOLOv8n, and YOLOv9n” the identification module

exactly distinguishes objects inside pictures. The

framework makes use of cutting-edge models for

characterization, such as "ResNet152,

EfficientNetB2, Exception, and Nonmobile," in

addition to a proprietary cross-breed model that

combines "DCNN and ResNet152 and an outfit

model that combines Exception and Nonmobile."

Using metrics such as "mean Average Precision

(map), F1-score, recall, accuracy, and precision,

“execution is entirely surveyed areas of strength for

ensuring across both recognition and characterization

errands.

3.1 Dataset Collection

MRI images which fall into both harmless and

threatening tumour classes - make up the dataset

utilized for brain tumour detection and arrangement.

Publically available clinical picture files

remembering the Brain Tumour Dataset for Kaggle

give the dataset and there are many named MRI scans

that are pre-handled to guarantee uniformity in size

and quality. A complete rationale for model training

is provided by the division of the data into several

Detection of Brain Tumors Using Advanced Image Processing and the Ensemble Model and YOLO Family

629

growth grades: "Grade 0 (no tumour), Grade I (small),

Grade II (medium), and Grade III (big)".

3.2 Pre-Processing

3.2.1 Classification

Augmenting Image Data: Enhancement of image data

for characterization refers to different methods

designed to work on the dataset and then increase

model guesswork. Re-scaling the picture to normalize

the size, shear changes to add minor mathematical

bends, zooming in to get better subtleties, and level

flips to duplicate a few survey points. Changing the

picture likewise ensures consistency in extents, so

empowering great preparation. These strategies

together increment the dataset and empower the

model to more readily oversee variances in genuine

visual data.

3.2.2 Detection

Image Processing: Image processing for detection

comprises in a few significant stages intended to

prepare the information for model surmising. To

normalize the picture for input, it initially gets

transformed into a mass item. The class is hence

determined; next comes announcing the jumping box

for detection. To simplify dealing with, the picture is

transformed into a NumPy array. The organization

layers are perused for stacking the pre-trained model

and result layers are separated. Added are the picture

and explanation records, which make an

interpretation of BGR to RGB, produce a veil, and

resize the image to fit the information particulars for

the model.

3.2.3 Data Augmentation

In detection, data augmentation alludes to strategies

that further develop the summing up limit of the

model. To start with, the picture is randomized to add

preparing changeability. Arbitrary pivots then, at that

point, are utilized to imitate different directions,

consequently directing the model to learn invariant

properties. To additionally fluctuate datasets, picture

changes including mutilation, interpretation, or

scaling are likewise finished. These expansion

methods ensure the model's capacity to productively

recognize objects from a few points, positions, and

circumstances, consequently fortifying its

presentation and strength in reasonable settings.

3.3 Algorithms

3.3.1 For Classification

DCNN: Deep elements from brain tumour pictures

are extricated utilizing DCNN, which gains muddled

examples and isolates harmless from cancer cases

along these lines empowering dependable tumour

grouping.

ResNet152: Reset 152 is utilized since it can

oversee deep neural networks, consequently working

on the limit of the model to learn and classify growth

photographs with higher accuracy by residual

learning methods.

EfficientNetB2: Utilizing its adaptable

engineering to deal with brain tumour pictures with

less boundaries and quicker training times,

EfficientNetB2 is utilized to amplify arrangement

accuracy while safeguarding effectiveness.

DCNN + ResNet152 - Hybrid Model: Meaning to

involve the two plans' assets for further developed

highlight extraction and more prominent accuracy in

distinguishing brain tumour pictures, the crossover

model mixes DCNN and ResNet 152.

Exception: Exception’s powerful convolutional

design assists the organization with extricating

undeniable level data from growth pictures, hence

giving improved characterization results to brain

tumour identification.

NasNetMobile: Especially accommodating for

brain tumour picture examination with restricted

assets, Nonmobile gives lightweight execution while

keeping up with extraordinary grouping accuracy,

thus empowering successful component extraction in

versatile settings.

Exception + NasNetMobile - Ensemble Model:

Consolidating “Exception with NasNetMobile”

permits the group model to take utilization of the two

organizations' advantages, subsequently further

developing arrangement accuracy by joining different

component separating powers from the two models.

3.3.2 For Detection

YoloV5x6: Through continuous item discovery

abilities, YOLOv5x6 gives rapid handling and precise

distinguishing proof of cancer regions, consequently

empowering identification of irregularities in brain

tumor images.

YoloV5s6: More modest tumor regions can be

successfully distinguished utilizing YOLOv5s6,

which ensures quicker execution for constant

applications and jelly extraordinary accuracy in

anomaly recognition.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

630

YoloV8: On account of its extended engineering

and component extraction strategies, YOLOv8 is

utilized for refined discovery occupations; it offers

higher accuracy and speed in growth finding,

especially in confounded pictures.

YoloV9: Coordinated for its "state-of- the-art

detection" abilities, YOLOv9 gives uncommon

accuracy in spotting brain tumor abnormalities with

few bogus up-sides, so ensuring steady outcomes for

use in centers.

The table 1 illustrate the performance

evaluation table for classification and table 2 shows

Performance Evaluation Table for Detection.

4 PERFORMANCE METRICS

Accuracy.

Accuracy 

























(1)

Precision.

Precision 













(2)

Recall.

Recall 













(3)

F1-Score.

F1 Score  2 ∗

  



∗ 1001 (4)

MAP.

"𝑚𝐴𝑃 





∑

𝐴𝑃







" (5)

Where,

MAP- mean average precision

AP- the “average precision

K-over all clients or searches

For Classification.

Table 1: Performance evaluation table for classification.

Model Accuracy Precision Recall F1-Score

EfficientNetB2 89.7 89.7 89.7 89.7

ResNet152 77.5 82.3 74.2 76.9

DCNN 81.1 55.0 93.1 67.9

Exception 85.4 86.9 83.0 84.3

NASNetMobile 92.9 93.1 93.0 93.0

For Detection:

Table 2: Performance evaluation table for detection.

Model Precision Recall mAP

YOLOV5s6 86.3 79.3 86.9

YOLOV5x6 73.3 60.7 81.7

YOLOV8 73.2 80.5 86.7

YOLOV9 84.3 78.9 78.6

5 RESULTS

The figure 2 shows Uploading an Input Image for Detection

and figure 3 shows Final Outcome. The figure 4 and 5

illustrate Uploading an Input Image for Classification and

Final outcome.

Figure 2: Uploading an input image for detection.

Detection of Brain Tumors Using Advanced Image Processing and the Ensemble Model and YOLO Family

631

Figure 3: Final outcome.

Figure 4: Uploading an input image for classification.

Figure 5: Final outcome.

6 CONCLUSIONS

The suggested approach for brain tumor differential

verification and programming demonstrates a notable

increase in accuracy and reliability by utilizing state-

of-the-art machine learning and deep learning

techniques. Through the integration of "deep

convolutional neural networks (DCNNs)" with

cutting-edge transfer learning models, the approach

successfully divides mental diseases into four

different grades: "Grade 0 (no tumor), Grade I (little

tumor), Grade II (medium-sized tumor), and Grade III

(big tumor)." Although the combination model of

"Exception and NasNetMobile" achieves exceptional

performance with "accuracy of 98.1%, precision of

98.3%, recall of 97.9%, and F1 score of 98.1%," the

DCNN and ResNet152 mixing model provides a

respectable standard for cancer order. With a recall of

78.9%, precision of 84.3%, and mean average

precision (mAP) of 78.9%, These discoveries show

how definitively the framework recognizes and

arranges cancers, which makes it a significant

instrument for clinical navigation. The intuitive point

of interaction in light of Flask ensures safe

confirmation and effortlessness of purpose, in this

manner overcoming any issues among innovation and

helpful medical services applications. Extending the

dataset to integrate a more shifted range of tumour

sorts and imaging settings will assist with working on

model speculation and hence the future degree of this

brain tumour detection and grouping technique.

Clinically, coordination with ongoing MRI or CT

scan information for live tumour identification and

classification could further develop handiness.

Counting “explainable artificial intelligence (XAI)”

approaches would likewise assist with expanding

process transparency for deciding. Further refining of

the YOLO demonstrates for quicker and more exact

distinguishing proof could additionally work on

continuous relevance in clinical circumstances.

Data set link: https://www.kaggle.com/datasets/

masoudnickparvar/brain-tumor-mri-dataset/data.

REFERENCES

A novel method for detecting and identifying brain tumors

using transfer learning. Applied Sciences,

Issue:12(Volume :11), Pg-5645. Ullah, Nissar., Khan,

Junaid. Aamjad., Khan, Mohammed. Siraj., Khan,

Walleha., Hassan, Ismail., Obayya, Mourice., Salama,

Alka. Sharma., 2022.

A comprehensive review of machine learning, hybrid deep

learning, and transfer learning techniques for Magnetic

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

632

Resonance Imaginging- based classification & segme-

ntation in brain tumor analysis. Multi-media Tools &

Appls, Pg-1-38. Das, Sumit., & Goswami, Rahul.

Sandesh. ,2024.

An advanced Brain tumor Detection and Classification

system in medical image processing utilizing a hybrid

deep CNN-Cov-19-Res-Net Transfer learning model.

Biomedical Signal Processing and Control, Kumar,

Kiran. Arun., Prasad, Aja. Y., & Metan, Jonnes.,2022.

An efficient hybrid transfer learning model for medical

diagnosis of brain tumors within a classification

framework. Diagnostics, Issue 12(Volume 10), pg

2541. Samee, Niakath. Abrar., Mahmoud, Nouman.

Fatehi., Atteia, Golab., Abdallah, Haleed. Amir.,

Alabdulhafith, Mohsi., Al-Gaashani, Moin. Sheikh., &

Muthanna, Multan. Sialkot. Abbatabad.,2022.

CNN transfer learning approach for automated brain tumor

detection. Medical and Biological Engineering and

Computing, Issue 61(volume 7), Pg-1821-1836.

Bairagi, Viraha. Kumar., Gumaste, Pratima., Rajput,

Solace. Hulka., & Chethan, Ketan. S.,2023.

Employing deep learning convolutional neural network

with transfer learning for brain tumor detection. Internl.

Journ. of Imaging Systems & Tech., Anjum, Sahay.,

Hussain, Lukkal., Ali, Meer., Monagi H. Alkinani., &

Duong, T. Quresihsi.,2022.

Enhancing the efficacy of various deep transfer learning-

based models in brain tumor detection from MR

images. IEEE Access, Volume 10, Pg34716-34730.

Asif, Suraigha., Yi, Wien., Ain, Quaan. Ushu., Hou,

Jin., Yi, Tia., & Si, Juan.,2022.

Enhancing brain tumor classification using a combination

of CNN & transfer learning. Knowledge-Based

Systems, 111981. İncir, R., & Bozkurt, F.,2024.

Evaluating deep transfer learning approaches for brain

tumor classification performance using MRI images.

Journ.of Healthcare Engg., Issue 2022(Volume 1), Pg-

3264367. Srinivas, Chinna., Kumar Singh, Nutan .

Parashar., Zakariah, Mushir., Porashar Alothaibi,

Yall.Aali., Shaukat, Khaled., Partibane, Bhilal., &

Awal, Hulaik.,2022.

Hybrid ‐ NET: Combining DenseNet169 and advanced

machine learning classifiers for improved brain tumor

diagnosis. Internl. Journ. of Imaging Sys. & Tech.,

Khan, Sohail. Ushu . Rixckie., Zhao, Muah., Asif,

Sheikh., & Chen, Xian.,2024.

Image-based brain tumor detection and comparison of

transfer learning methods with 3-layer Convolutional

Neural Network. Khaliki, M. Z., & Muhammet Sinan

Başarslan.,2024.

Improving brain tumor classification through an extensive

analysis of transfer learning techniques & model

efficiency using Magnetic Resinance Imaging datasets.

Shamshad, Numaid., Sarwr, Delloitte., Almogren,

Archin., Saleem Khan, Khaled., Munawar, Ali.,

Rehman, Aamir. Umar., & Bharany, Shahid. ,2024.

Utilizing hybrid deep learning models for brain tumor

detection and classification. Scientific Reports,

Volume13(Issue - 1), Pg-23029. Babu Vimala, Balak.,

Srinivasan, Shohesh., Mathivanan, Sisir. Kumar.,

Mahalakshmi, Jayagopal, Prakash., & Dalu, Gopal.

Timor. ,2023.

Utilizing advanced machine learning techniques and

knowledge transfer for precise brain tumor

identification. Mathivanan, S. Kumar., Sonaimuthu,

Srinivasa., Murugesan, Shanmukha., Rajadurai,

Hitesh., Shivahare, Bolan. Duleep., & Shah,

Mohammaed. Aziz.,2024.

Detection of Brain Tumors Using Advanced Image Processing and the Ensemble Model and YOLO Family

633