Brain Tumor Detection of MRI Images Using CNN Implemented on

VGG16 Based Architecture

Chudaman Sukte, Tarun Tiwari, Gautam Nimase and Sandesh Mahajan

Dept. Information Technology Vishwakarma Institute of Information Technology Pune, India

Keywords: Brain Tumor Detection, MRI, CNNs, VGG16 Architecture, Transfer Learning.

Abstract: From a medical perspective, brain cancer can be considered one of the most lethal diseases due to the damage

to major blood vessels and the increased risk of death. Therefore, early and accurate diagnosis is important

for the best treatment of the disease. In this paper, we describe a new method for automatic problem detection

based on the VGG16 neural network, which recognizes the deep structure and good image distribution. Our

model involves the enhancement of MRI scan images and then the classification of images into tumor and

non-tumor using transformations with VGG16. We build models that achieve satisfactory accuracy,

sensitivity and specificity using large-scale MRI images and their annotations. Our results show that the

VGG16 mathematical model can assist radiologists in brain diagnosis and make brain diagnosis more efficient

and reliable. Additionally, we provide an overview of the possibilities of deep learning in modern medicine

and the prospects for the development of medical imaging.

1 INTRODUCTION

Many patients around the world are going through an

end-of-plan analysis that glaringly affects survival

costs.

Brain tumors are life-threatening and cause

globally unique neurological problems with a high

mortality rate. Early will is critical to convincing

treatment and advanced survival fees. MRI is the

standard non-invasive strategy used to distinguish

brain tumors, advertising and marketing stupid and

crude images of mind tissue. In any case, post-

examination of MRI filters is time-consuming and

prone to human error, especially for large data sets or

subtle abnormalities. Millions of mind tumor

sufferers regularly face delayed willpower due to the

need for assets or radiologists, especially in

underserved areas such as provincial parts of Asia

[(Siar, and Mohammad 2019).

The early end is huge because appropriate

treatment increases overall survival costs and

improves knowledge outcomes. Orientation

examination of the appearance of MRI, which is a

modern general strategy for the detection of brain

tumors, is unfortunately not lengthy, but is also prone

to human error, especially when there is a large

amount of information, or inconspicuous anomalies

that lend themselves to being ignored. This occurs

around a delayed or incorrect examination, which can

extremely affect the quality of care and prognosis

(Siar, and Mohammad 2019), (Deshmukh, and

Bendre 2024).

Our goal is to create a proficient mind tumor

discovery machine utilizing VGG16 deep mastering

reveal to return appropriate determination.

Our proposed framework, built on VGG16

engineering and using business knowledge, considers

handling these challenging situations by mechanizing

the discovery of brain tumors in MRI filters. The

intention of the program is to catch really the smallest

and most inconspicuous tumors and ensure that no

key element is missed. By joint mechanization, we

reduce the burden on radiologists, speed up

demonstrative processing and increase the accuracy

of symptoms. This framework will turn out to be

particularly important in places with limited access to

filling specialists, bridging the hole in healthcare

administration and saving lives with timely and

accurate mind tumors.Utilizing the advanced and

profound tactics that VGG16 encompasses, we point

out the creation of a robotic, remedial, and accessible

framework for mind tumor location. This gadget will

assist healthcare professionals in faster and

remarkably more accurate examinations, ultimately

improving chronic outcomes and taking care of the

Sukte, C., Tiwari, T., Nimase, G. and Mahajan, S.

Brain Tumor Detection of MRI Images Using CNN Implemented on VGG16 Based Architecture.

DOI: 10.5220/0013633900004664

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

In Proceedings of the 3rd Inter national Conference on Futuristic Technology (INCOFT 2025) - Volume 3, pages 627-635

ISBN: 978-989-758-763-4

627

global health hole (Ramprakash et al. 2024),

(Rehman, Amjad, et al. 2023).

Our aim is to make brain tumor localization faster,

more robust and manageable for everyone, regardless

of geological or financial constraints, thereby

contributing to higher survival fee and much higher

knowledge outcomes across the sector.

This project will contribute to the research and

society in a remarkable way:

1. Advances in Therapeutic Imaging: By

engineering VGG16 and acquiring

alternative knowledge, it includes expansion

into the evolving framework of deep study

techniques for therapeutic imaging,

particularly brain tumor discovery. It

explores how pre-trained models can be best

tuned for specialized tasks, pushing the

boundaries of existing strategies in radiology

and AI-assisted prognostication.

2. Advances in demonstrative accuracy: Our

images highlight how thorough mastery can

distinguish subtle deviations from the norm

in MRI filters, likely bypassing traditional

guided investigations. This will open up a

modern exploration of approximately paths

in growing larger contemporary models that

could deal with complicated healing photos.

3. Benchmarking and replicability: By sharing

and approximately derived techniques, we

provide a gadget that can help improve fate

analysts, cultivate collaboration and progress

within the subject of automated recovery

diagnostics.

This paper is prepared as follows: phase II affords

the literature overview and associated works in mind

tumor detection using deep learning. segment III

describes the methodology & the VGG16 model

utilized in our study. segment IV discusses the results

received from the experimental assessment, even as

phase V interprets the findings and highlights key

insights. In the end, phase VI concludes the paper

with a precise capability for future paintings (Preetha,

Jasmine et al. 2024).

2 LITERATURE SURVEY

Table 1: Literature survey

Author(

Focus of

the Paper

Key Points

Covera

Methodology(s)

Used

Al-

Ayyoub

et al

Machine

Learning

for Brain

Tumour

Detection

Comparing

the

performanc

e of

different

machine

learning

algorithms

for brain

tumour

detection

using MRI

ima

es.

Image

Processing:

Conversion of

RGB images to

greyscale.

Machine

Learning

Algorithms:

ANN, Tree J48,

Naive Bayes,

and LazyIBk.

Hemant

h et al.

Brain

Tumour

Detection

using

Machine

Learning

Proposing

a brain

tumour

detection

system

using

machine

learnin

The specific

machine

learning

approach used

is not specified.

Shishir

et al

Brain

Tumour

Detection

using

CNN

Using

Convolutio

nal Neural

Networks

(CNN) for

brain

tumour

detection.

Convolutional

Neural

Networks

(CNN)

Nandpu

ru et al

MRI Brain

Cancer

Classificati

Classifying

brain

cancers

from MRI

data.

Support Vector

Machine

(SVM)

Chandr

a &

Kolasan

Brain

Tumour

Detection

Detecting

brain

tumours.

Genetic

Algorithm

Varuna

Shree

T.N.R.

Kuma

Brain

Tumor

MRI

Image

Identificati

on and

Classificati

Identifying

and

classifying

brain

tumour

MRI

images

using

feature

extraction.

Discrete

Wavelet

Transform

(DWT)Probabil

istic Neural

Network

Younis

et al

Brain

Tumour

Analysis

Analysing

brain

tumours.

Deep Learning

Irmak Brain

Tumour

MRI

Image

Multi-

classificati

Multi-class

classificati

on of brain

tumour

MRI

images.

Deep

Convolutional

Neural Network

Jia &

Chen

Brain

Tumour

Identifying

and

Deep Learning

INCOFT 2025 - International Conference on Futuristic Technology

628

Identificati

on and

Classificati

classifying

brain

tumours

from MRI

images.

Byale

et al.

Brain

Tumour

Segmentati

on and

Classificati

Automatic

segmentati

on and

classificati

on of brain

tumours.

Machine

Learning

Alquda

h et al

Brain

Tumour

Classificati

Technique

Comparing

different

brain

tumour

classificati

techniques.

Deep Learning

Amin et

Brain

Tumour

Detection

and

Classificati

Surveying

methods

for brain

tumour

detection

and

classificati

on.

Machine

Learning

The subject of brain tumour detection and

classification using MRI images has benefited greatly

from recent developments in machine learning (ML).

Al-Ayyoub et al. investigate a number of machine

learning techniques, such as Artificial Neural

Networks (ANN), Decision Trees (J48), Naive

Bayes, and LazyIBk. They show that once MRI data

are converted to greyscale format, the performance of

various algorithms can differ for cancer identification

tasks. Despite without naming the exact approach,

Hemanth et al. offer a broad ML-based detection

framework. Shishir et al. demonstrate the efficacy of

deep learning models in managing intricate image-

based tasks by using Convolutional Neural Networks

(CNN) specifically to MRI images for more complex

models Lin, et al. 2023), (Liu, et al. 2024).

3 METHODOLOGY

3.1 Dataset Collection

In our study, we used the brain tumor MRI image

dataset from Kaggle, which contains 3064 MRI

images containing one of the 17 unique features that

define brain tumors. The image is then converted to

224 x 224 pixels, which is required to enter the

VGG16 standard. The dataset is divided into training

set (70%), validation set (15%), and test set (15%).

Figure 1: Methodology

3.2 Data Augmentation

RandomHorizontalFlip(p=0.5):which behaves

the same as above horizontally flipping with 50%

probability i.e., augments model to generalize against

flipped objects also but meantime we wanted our

model to recognize the view irrespective of its

position and hence there was no need of rotation.

RandomVerticalFlip(p=0.5): It will randomly

paste Image vertically with 50% Acknowledgment

and Will help to add more diversity into the dataset

RandomRotation(degrees=15) : Application of

this performs a random rotation to our image with an

angle within 15 degrees which can sometimes be

helpful in minor rotations and real-world distortions

of an image.

ColorJitter(brightness=0.2, contrast=0.2):

randomly change the brightness and contrast of the

image to simulate lighting effects

RandomResizedCrop(size=(224, 224),

scale=(0.8, 1.0)): This will randomly crop image to

the specified dimension then resize it to that size but

keep the whole area from the original image in which

it was cropped at least scaled by 80% and scaled up

to 100%.

ToTensor(): This will convert the image into a

tensor that your model can take as an input.

Normalize(mean=[0.485, 0.456, 0.406],

std=[0.229, 0.224, 0.225]): Normalize an image to

the mean and standard deviation specified by

ImageNet-trained models

This augmentation helps the model to learn more

complex images in order to gain high accuracy for

over 17 classes .

3.3 Model Architecture

The convolutional neural network model originating

from the brain detection VGG16 framework has 5

Brain Tumor Detection of MRI Images Using CNN Implemented on VGG16 Based Architecture

629

groups, each group has a series of convolutional

processes, and the most common process is used for

subtraction after batch normalization. The number of

filters in the first block is as many as 64, and the filters

in the deep blocks are 512, which can detect low and

high levels. This model is separated by three

connections of all layers, and then RELU activation

and batch recovery. The urine process to distribute the

MRI section as much as possible.

Clearly! The following is an explanation of each

layer in the VGG16-based model and how it fits into

the steps of brain tumor detection and classification;

Figure 2: Model Architecture

3.3.1 Layer descriptions

1.Convolution layer: (feature extraction)

nn.Conv2d layers : The current layer extracts

features from the input image by creating a set of

filters on it. These filters look for specific local

patterns such as edges, textures, and other simple

image elements. As we move through the network,

these filters learn to pick up more sophisticated

patterns, such as features that are specific to tumors.

nn. BatchNorm2d: Normalization for each batch

after convolution. It normalizes its performance so

that the next layer has more stable values, which helps

with faster training and resistance to weight

initialization.

Activation function (F.relu): ReLU is a building

block = applied between each pair of convolutions. It

introduces non-linearity, making the network model

complex functions.

2. Pooling layers (reducing dimensions)

nn. MaxPool2d layers: Max Pool2d layers

reduce the spatial dimensions of feature maps, reduce

computation, and focus on critical features. This

increases performance by making the model

computationally efficient and avoids overfitting and

suppression of less important details.

3.Classification — Fully connected layers:

nn. Flatten Layer : Merges the output of the last

convolution block so that it is fed into fully connected

layers.

nn. Linear layers: These layers are the classifier.

They take the features learned by the convolutional

layers and map them to the output classes, which in

this case, are the different kinds of brain tumor. The

first two fully connected layers, that is, the fc1 and

fc2, help the model learn complex patterns &

combine the different features. Finally, the features

are mapped to the output classes with the last fully

connected layer, fc3.

nn. BatchNorm1d layers: for fully connected

layers that stabilize and accelerate learning with batch

normalization

4. Output layer:

The final output layer (the attention layer) predicts

the label relative to the brain tumor class. Additional

probabilities can be derived based on the outputs by

using them depending on the loss function that was

used during training (e.g. CrossEntropyLoss).

Role Summary:

Feature extraction (Conv2d + ReLU +

BatchNorm2d): Convolutional operations capture a

hierarchy of spatial features starting from edges and

finally capture tumor-centered features.

Dimensional reduction (MaxPool2d):

Maximum pool layers gradually reduce dimensions,

which helps the network to be more compact in terms

of depth, and also helps to keep only the necessary

components.

Classification (Flatten + Linear +

BatchNorm1d): These are computational

components involved in deriving associations

between classes mapped by brain tumors and features

obtained from convolutional layers.

The above feature extraction followed by

classification was constructed for brain tumor

detection and classification using the VGG16 feature

representation(Neamah, Karrar, et al. 2023)

5. Loss and optimizer:

When constructing the display, we don't forget the

truth that CrossEntropyLoss() is the maximum

counseled misfortune work, for the reason that that is

a multi-magnificence classification demonstrated.

This misfortune is precious in deciding how remote

the expected yield is from the real taking a toll, and

this makes a distinction in the widespread mastering

of the version.

The optimizer applied for our show is

Adam(version.Parameters(),lr=zero.001,

weight_decay=1e-4, betas=(0.Nine, zero.999)):

INCOFT 2025 - International Conference on Futuristic Technology

630

lr (learning price=zero.001): Controls the step

measure at each cycle whilst shifting to the least

misfortune paintings. A little esteem like zero.001

makes the gaining knowledge extra solid.

Weight_decay=1e-four: An administrative term

that makes a distinction to avoid overfitting by

penalizing expansive weights.

Betas=(0.9, 0.999): Coefficients applied to

calculate the running midpoints of the slope and its

square:

beta1=zero.9: Rate of decrease for the first of all

moment (normal of gradients).

Beta2=0.999: Decay fee for the instant second

(uncentered fluctuation of gradients).

6.Performance Metrics:

Accuracy:

3.3.2 Mathematical definition

Description: Exactness measures the volume of

modified expectations made by means of the display

out of the upload as much as quantity of events. It is

a not unusual diploma of show execution, but may not

be pretty enlightening if the records set is choppy (e.g.

whilst one route is more go to than others) (Liu, Min,

et al. 2024).

3.3.3 Confusion Matrix

Description: The disarray network offers a nitty

gritty breakdown of display expectancies in

comparison to real names, appearing actual positives,

wrong positives, proper negatives, and unfaithful

negatives for each class.

The disarray network permits you to visualize

execution over various instructions of mind tumors,

creating a distinction to determine which sorts are

bewildered by using the exhibit. This can also lead to

changes in information series or exhibit structure

(Liu, et al. 2024).

3.3.4 Mathematical representation

True Positive (TP): Accurately anticipated high-

quality cases.

True Negative (TN): Accurately anticipated

terrible cases.

False fine (FP): Inaccurately expected nice

instances.

False bad (FN): Erroneously expected negative

instances.

3.3.5 Structural Closeness File (SSIM)

Description: SSIM assesses the likeness among

two photographs and offers a picture quality metric.

It takes brightness, differentiate and surface into

account.

SSIM may be applied to assess the exceptional of

the yield images produced by your show, in particular

in case you utilize photograph amplification methods.

It can provide assistance to decide whether or not the

show jam crucial fundamental factors of interest in

the pix.

3.3.6 F1 Score

Description: The F1 rating is a consonant cruel of

exactness and evaluation and gives a adjust among

the two

The F1 rating is in particular treasured in

restorative programs along with mind tumor class,

wherein wrong poor comes about may have proper

outcomes. A tall F1 score implies your show has a

brilliant modification between exactness and

evaluation over numerous instructions.

Brain Tumor Detection of MRI Images Using CNN Implemented on VGG16 Based Architecture

631

3.3.7 Calibration curve

Description: Calibration curves show how well

the predicted probability matches the correct results

and distinguish whether the demonstration is well

calibrated or not.

The calibration curve plots the predicted

probabilities as opposed to the distribution of positive

values (Pattanaik, Sudeshna, et al. 2024).

4 RESULTS AND DISCUSSION

The sequel is going to demonstrate the display design

that has been in the making for a hundred and fifty for

a long time, using the NVIDIA Tesla T4 GPU in the

Google Colab. Robust GPU computing manipulation

with a potential of 64GB/s of reminiscence transfer

and 16GB of VRAM allowed the show to effectively

process a large amount of statistics for more than an

hour and achieve results after a characteristic

execution. Metrics (Preetha, Jasmine et al. 2024).

4.1 Evaluation of performance metrics:

4.1.1 Accuracy

It measures charge as it should be categorized by

occasion, which usually proves to reveal performance

(Rehman, Amjad, et al. 2023).

Figure 3: Accuracy

4.1.2 Specificity

It quantifies the ability of the version to appropriately

discriminate against non-tumor cases, demonstrating

viability in minimizing false positives.

Figure 4: Specificity

4.1.3 F1 score

It combines precision and insight to provide an

adjusted degree of type execution, with high values

showing much better results (Preetha, Jasmine et al.

2024).

Figure 5: F1 Score

INCOFT 2025 - International Conference on Futuristic Technology

632

4.1.4 Confusion Matrix

Visualizes true positives, false positives, authentic

negatives and false negatives for each lesson,

differentiates understanding of misclassification

patter(Neamah, Karrar, et al. 2023)

Figure 6: Confusion Matrix

4.1.5 Calibration curve

It evaluates the arrangement of expected probabilities

with actual consequences; focus close to the tilted

exhibit large calibration (Wageh, et al. 2024).

Figure 7: Calibration Curve

4.1.6 Basic Similarity Record (SSIM)

It assesses the similarity between exact and expected

images; higher values indicate advanced preservation

of additional lights.

Figure 8: Similarity Record (SSIM)

4.2 Discussion

4.2.1 Key Findings

Model execution: Highlight the model's ability to

accurately classify brain tumors using CNN

engineering and highlight the high accuracy and

typical fit of your method

Metrics Achievements: Seemingly noteworthy,

achieved with metrics such as accuracy, F1 rating,

SSIM, and calibration bends, providing accurate

reports on model accuracy and unwaveringly

exceptional in real global programs.

Optimization Strategy: Explore the impact of

using optimizers like Adam with properly tuned

hyperparameters (gain knowledge of charge, mass,

beta) for incremental merging and overall

performance.

Computing Prowess: Be aware that it is part of

using a GPU (NVIDIA Tesla T4) to speed up

preparation and boost successful large-scale

recording preparation.

4.2.2 Model performance

Strengths:

1. High preparation accuracy (ninety-

five.06%): The display learned to effectively

understand the designs within the preparation

facts, showing that the design is properly

perfect for the task and data set. This high

accuracy reflects excellent study skills.

2. Strong class performance: F1 scores for

several classes are greater than 0.90 (e.g. class

zero: 0.9040, class 1: zero.9107, course four:

0.9508), indicating the adjusted accuracy and

ranking for these classes. Publish behaves quite

correctly when looking ahead to these classes.

3. Balanced F1 scores in most classes: Classes

that include nine (0.9444) and 11 (0.9310)

appear to be reliable and stable performers,

illustrating the version's potential to generalize

well across categories (Neamah, Karrar, et al.

2023)

Brain Tumor Detection of MRI Images Using CNN Implemented on VGG16 Based Architecture

633

weaknesses:

1. Testing Accuracy (86.46%): The difference

between preparation and testing accuracy

(approximately 9%) suggests reassembly. The

show probably learned to draft stats by rote,

causing it to underperform on subtle control

information.

2. Characteristic Biased Class: Lower F1 scores

in instructions that contain 16 (zero.7234) and 5

(0.7642) may be the end result of lesson

imbalance or insufficient statistical tests for

these categories, causing the display to struggle

with accurate predictions.

3. Higher trial calamity (0.404): The comparison

of prepared calamity (-0.153) and reported

calamity seems to be trying to generalize. A

high calamity harbinger regularly focuses on

issues with shouting or hazing in connection

with an impending calamity (Lin, et al. 2023),

(Pattanaik, Sudeshna, et al. 2024).

Possible reasons:

1. Training Time and Complexity: The sample

took 2 hours to compile over a hundred and fifty

on an Nvidia Tesla T4 GPU. The delayed setup

time suggests that the show may be complicated

and more tuning (e.g. regularization techniques)

may also relieve overfitting.

2. Imbalanced information: A few lessons are

likely underrepresented in the dataset, causing

the display to behave worse than the views

considered in the F1 score, and distort grid

inconsistencies.

4.2.3 Challenges They Face

Over the course of my show, I've done a few

challenges that I've won through inspection and

specialized upgrades.

Low accuracy without augmentation:

• Initial preparation without expanding the

records introduced in about lousy accuracy.

• The display tried to generalize due to the

limited variability of the data set.

• Information dissemination methods

explored and updated, with advances

showing robustness.

Extended preparation time:

• Exercising on a nearby machine has turned

into a waste and a waste of time.

• It used an NVIDIA Tesla T4 GPU from

Google Colab, which reduced preparation

time to honesty by hours.

Insufficient ranking metric:

• The initial performance evaluation required

accurate evaluation metrics.

• Explore advanced metrics (eg SSIM, F1

score, specificity, calibration curve) for

comprehensive performance evaluation.

• The implementation of these measurements

made it possible to correctly recognize the

qualities and shortcomings of the program

(Liu, et al. 2024), (Pattanaik, Sudeshna, et

al. 2024).

5 CONCLUSIONS

In this paper, we faced many challenges in building a

strong classification sample. Initially, our show

struggled with execution due to missing information

that occurred with moo accuracy when using raw

information without augmentation. After viewing

several investigative documents, we created a flood

of information that overall moved forward and

showed generalization and accuracy. The long

preparation time was another challenge that we

overcame by using the Google Colab GPU (NVIDIA

Tesla T4), reducing the total preparation time to 2

hours (Liu, et al. 2024).

Moreover, despite the fact that the initial

demonstration yielded great accuracy, the need for in-

depth implementation measurements limited the

investigation. To address this, we unified measures

such as F1 score, approach record (SSIM),

specificity, and calibration bend, driven by a paper

reference query, to gain a more comprehensive

experience of the model's qualities and shortcomings.

These extraordinary measurements revealed areas

where the demonstration exceeded expectations and

where progress could be made, such as the tendency

for over-fitting and course imbalance (Rehman,

Amjad, et al. 2023).

We also tested with the VGG16 design, including

unused features that contributed to significant

improvements in highlighting extractions and general

classification performance. By combining advanced

evaluation metrics and combining well-known deep

learning with enhancements, this reasoning lays the

groundwork for future optimization and progress in

image classification matching (Liu, et al. 2024),

(Pattanaik, Sudeshna, et al. 2024).

INCOFT 2025 - International Conference on Futuristic Technology

634

REFERENCES

Siar, Masoumeh, and Mohammad Teshnehlab. "Brain

tumor detection using deep neural network and machine

learning algorithm." 2019 9th international conference

on computer and knowledge engineering (ICCKE).

IEEE, 2019.

Deshmukh, Aishwarya V., and M. R. Bendre. "Multiclass

Brain Tumor Detection using Deep Transfer Learning."

2024 MIT Art, Design and Technology School of

Computing International Conference

(MITADTSoCiCon). IEEE, 2024

Ramprakash, B., et al. "Comparative Analysis of Deep

Learning and Machine Learning for Detection and

Classification of Brain Tumors at Multiple Stages."

2024 Second International Conference on Emerging

Trends in Information Technology and Engineering

(ICETITE). IEEE, 2024..

Rehman, Amjad, et al. "Detection of Lungs Tumors in CT

Scan Images Using Convolutional Neural Networks."

IEEE/ACM Transactions on Computational Biology

and Bioinformatics 21.4 (2023): 769-777

Preetha, R., M. Jasmine Pemeena Priyadarsini, and J. S.

Nisha. "Automated Brain Tumor Detection from

Magnetic Resonance Images Using Fine-Tuned

EfficientNet-B4 Convolutional Neural Network." IEEE

Access (2024).

Wageh, Mohamed, et al. "Brain Tumor Detection Based on

Deep Features Concatenation and Machine Learning

Classifiers With Genetic Selection." IEEE Access

(2024).

Lin, Xiaoyi, et al. "Improving tumor classification by

reusing self-predicted segmentation of medical images

as guiding knowledge." IEEE Journal of Biomedical

and Health Informatics (2023)

Liu, Huabing, et al. "Multimodal Brain Tumor

Segmentation Boosted by Monomodal Normal Brain

Images." IEEE Transactions on Image Processing

(2024).

Neamah, Karrar, et al. "Brain Tumor Classification and

Detection Based DL Models: A Systematic Review."

IEEE Access (2023).

Asiri, Abdullah A., et al. "Optimized Brain Tumor

Detection: A Dual-Module Approach for MRI Image

Enhancement and Tumor Classification." IEEE Access

12 (2024): 42868-42887.

Liu, Min, et al. "Brain Image Segmentation for Ultrascale

Neuron Reconstruction via an Adaptive Dual-Task

Learning Network." IEEE transactions on medical

imaging (2024).

Pattanaik, Sudeshna, et al. "A MIoT Framework of

Consumer Technology for Medical Diseases

Prediction." IEEE Transactions on Consumer

Electronics (2024)

Brain Tumor Detection of MRI Images Using CNN Implemented on VGG16 Based Architecture

635