Journal of Theoretical and Applied Information Technology

31st August 2024. Vol.102. No. 16

© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

OPTIMIZED DEEP LEARNING FRAMEWORK FOR BRAIN

TUMOR DETECTION AND CLASSIFICATION USING
HYBRID VISUAL GEOMETRY GROUP-16 WITH REDUCED
WEIGHTS VIA BUTTERFLY OPTIMIZATION
1,*
RAMYA NIMMAGADDA, 2Dr. P. KALPANA DEVI
1
Electronics and Communication Engineering (ECE)
Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology
Chennai, India
2
Electronicts and Communication Engineering (ECE)
Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology
Chennai, India

Email: [email protected], [email protected]

ABSTRACT
Finding and classifying brain tumors are important parts of medical image analysis that need
advanced deep-learning methods and optimization algorithms. Recognizing the urgent need for accurate
methods in brain tumor diagnosis, we present a comprehensive approach integrating various stages, including
data preprocessing. In this preprocessing phase, we employ techniques like aspect ratio normalization and
resizing to form a standardized dataset. By standardizing image dimensions, we aim to improve subsequent
processes like feature extraction and segmentation, reducing potential distortions. The suggested model is
made by using Convolutional Neural Networks (CNN) to find patterns and traits that make tumor and non-
tumor areas different from each other. To overcome the intricate sections and fine textures during down-
sampling, the proposed model is hybridized with U-Net architecture which gives accurate and robust results
of 98%. Furthermore, the Dice coefficient is measured using Intersection Over Union (IOU) to ensure
whether it is robust to class imbalance. This shows an intuitive interpretation, with higher values of 0.83 and
0.9 indicating strong and better segmentation performance. The model is further developed with VGG-16 to
classify the tumor grades. In terms of accurately segmenting the tumor grades, the learnt relevant
characteristics that are derived from the segmented tumor photos provide a 73% level of satisfaction. In
order to overcome the complexity and over-fitting problems, the Butterfly Optimization algorithm is
hybridized with VGG-16 which gives an enhanced output in classifying the grades. The proposed model
outperforms other Machine Learning (ML) and Deep Learning (DL) methods in tumor and non-tumor
identification and categorization with 99.99% accuracy. To further evaluate the suggested model's
performance, mobility, and energy economy, it is also implemented in JETSON Orin hardware.

Keywords-Deep Learning, Convolutional Neural Networks (CNNs), U-Net, VGG-16, Butterfly Optimization
Algorithm.

1. INTRODUCTION planning therapies for brain tumors [2]. The scope

of this research lies in recognizing the pressing need
The battle against brain tumors has reached for precision and efficiency in medical imaging
a critical juncture, where innovation in medical analysis to bolster brain tumor diagnostics and
imaging analysis stands as a beacon of hope [1]. The therapeutic interventions [4]. Traditional
intricate challenge of accurately detecting and methodologies, hampered by their limited accuracy
classifying these malignancies demands a paradigm and efficacy, fall short in meeting the demands of
shift in technological solutions. Our research reveals modern healthcare. Thus, the crux of our endeavor
a breakthrough framework that blends state-of-the- lies in surmounting these challenges through the
art deep learning methods with metaheuristic development of a holistic pipeline encompassing
optimization algorithms; this framework has the data analysis, pre-processing, segmentation, feature
potential to alter, when it comes to identifying and extraction, optimization, and classification phases.

6258
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

Our framework represents a pioneering reducing potential distortions. In order to

effort to amalgamate Convolutional Neural differentiate between tumor and non-tumor areas,
Networks (CNNs) with U-Net for meticulous our model uses Convolutional Neural Networks
segmentation and to integrate the formidable VGG- (CNNs) to detect complex patterns. To overcome
16 architecture with Butterfly Optimization for downsampling challenges, we hybridize our model
optimized feature extraction [3]. Brain tumor with the U-Net architecture, achieving a robust
identification and categorization reach new heights segmentation performance of 98%. We evaluate
of accuracy and efficiency because to the domino segmentation robustness using the Dice coefficient,
effect of these components working together. ensuring resilience to class imbalances. Expanding
Central to our approach is the concept of federated our research, we classify tumor grades with the
learning, a decentralized paradigm that respects data VGG-16 architecture, achieving 73% accuracy. To
privacy while harnessing the collective intelligence mitigate complexities and overfitting, we integrate
embedded within disparate datasets. Leveraging the the Butterfly Optimization algorithm, enhancing
Federated Averaging (FedAvg) algorithm [1], our classification performance.
framework trains on decentralized data sources Through rigorous evaluation, our model achieves an
without compromising sensitive information—a unparalleled efficiency of 99.99%, outperforming
pivotal breakthrough in medical AI. Drawing from a existing methodologies. Deploying it on JETSON
rich tapestry of references, our methodology is Orin hardware validates its performance and
fortified by the insights gleaned from recent portability, promising enhanced patient care
advancements in brain tumor classification and globally.
segmentation. Hasan et al. underscores the
transformative potential of transfer learning in The contributions of the article are as follows:
automating multi-class brain tumor classification, ● Introduction of novel preprocessing
paving the way for enhanced treatment strategies [2]. techniques for standardized dataset
In the same way, Gamal et al. argue that 3D U-Net formation.
can automatically separate brain tumors, which ● Integration of CNNs with U-Net
would start a new age of precision medicine [4]. architecture for robust segmentation.
Our system is proof of how powerful ● Rigorous evaluation using Dice coefficient
optimization methods can be for making deep and VGG-16 for tumor grade classification.
learning models work betterTo make brain tumor ● Enhancement of classification performance
segmentation more accurate, Ramírez et al. talk through Butterfly Optimization algorithm.
about how to use deep learning to make variational Organization of the paper:
models better [3]. Dolaat et al. showed how Following this, the paper has a
cooperative learning can help solve problems caused comprehensive methodology part that provides a
by uneven medical picture datasets. To move description of the framework for the detection and
healthcare AI forward [5], it's important for people categorization of brain tumors. Other writers who
to work together and share their thoughts. The results have conducted research on the segmentation and
of our study are a huge step forward in the search for categorization of brain tumors are covered in part II
better and faster ways to find and diagnose brain of this article. Following this, in section III, a
tumors. We are almost ready to start a new era of discussion is held on the description of each phase of
precision medicine where every patient gets the care, the process. These phases include data collecting,
they deserve by using the combined knowledge of preprocessing, segmentation, feature extraction,
deep learning and optimization algorithms. optimization, and classification. Following that, the
section on the outcomes of the experiment offers the
1.1 Motivation empirical findings that were gained from the
Brain tumor detection and classification are implementation of the suggested framework. These
paramount in medical imaging, demanding precise findings are complemented by visualizations and
diagnostic methods to avoid misdiagnoses and comparison analyses with methodologies that are
delays in treatment. Our research addresses this gap already in use. There is a detailed analysis of the data
by integrating advanced deep-learning techniques that is presented in the discussion section. This
and optimization algorithms. We prioritize data analysis addresses the ramifications, limits, and
preprocessing to establish a standardized dataset, potential future research areas. In conclusion, the
employing techniques like aspect ratio normalization conclusion provides a summary of the most
and resizing. This ensures uniform image important findings, reiterating the relevance of the
dimensions, streamlining subsequent processes and framework that was provided, and encourages more

6259
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

collaboration and creativity in the use of deep instructions. In 2020, Bhanothu et al. [9] introduced
learning to address difficulties in the healthcare a deep convolutional network to classify brain
industry. tumors in MRI scans. MRI scans are essential for
tumor diagnosis, therefore the scientists created
2. RELATED WORKS Faster R-CNN deep learning to automate tumor
identification and localization. It designates tumor
Many deep learning experiments have occurrence areas using Region Proposal Network
targeted brain tumor detection and classification. (RPN), accelerating review and reducing errors.
Dipu et al. [6] used YOLOv5 for object Regional proposal and categorization of glioma,
identification and FastAi for classification to achieve meningioma, and pituitary tumors utilizing VGG-16
good MRI scan accuracy (2021). To increase architecture. Evaluation results showed promising
detection accuracy and resilience, better CNN accuracy ratings of 75.18% for glioma, 89.45% for
architectures for tumor segmentation (2021), hybrid meningioma, and 68.18% for pituitary tumors. The
CNN-RNN models (2021), and ensemble learning mean average precision of automated brain tumor
(2021) have been investigated. Deep learning may detection and classification was 77.60% across all
change brain tumor diagnosis and therapy by classes.
enabling early detection and planning. In 2022, Abirami et al. [11] used
Sankaranarayaanan et al. [7] presented a VGG-16 Generative Adversarial Networks (GANs) to fuse
solution for brain tumor identification and PET and MRI images. They stressed the necessity of
classification at the 2023 International Conference multimodal picture fusion in medical applications to
on AI and Knowledge Discovery in Concurrent improve visual content and medical analysis. They
Engineering. As big data grows in medicine, the noted that multimodal fusion helps doctors view
authors stressed the importance of data analysis and hard and soft tissue, especially in brain imaging
mining on tumor prediction, monitoring, diagnosis, where tumor segmentation is critical. The suggested
and therapy. Brain tumors are aggressive and have method enhances tumor localization accuracy and
low survival rates, thus better diagnostic and diagnostic speed by fusing PET and MRI data. PET
treatment methods are needed. Deep learning, gives functional data while MRI provides
particularly CNNs, was used to solve brain cancer anatomical data to characterize brain cancers. The
diagnosis and treatment issues. Federated Learning proposed GAN-based fusion model blends different
(FL) was used to improve privacy and scalability in modalities to provide fused pictures for medical
centralized data collecting. The study optimized analysis and surgery planning. GAN-based model
model parameters for MRI tumor identification fared well with 0.8551 structural similarity and
using the VGG-16 architecture for brain cancer 2.8059 mutual information. Multimodal medical
detection and a CNN model framework. image fusion enhances brain tumor identification
Experimental findings showed that the suggested and treatment planning in this research. Haq et al.
approach outperformed traditional techniques with [12] suggested DACBT, a deep learning brain tumor
92% accuracy. This research advances brain tumor classification approach using MRI data in IoT
identification and shows how deep learning may healthcare, in 2022. Deep learning-based brain
improve healthcare. tumor classification was proposed to address the
Younis et al. [8] combined deep learning limitations of artificial diagnostic systems for brain
and ensembling learning to show that AI and neural cancer diagnosis in IoT-healthcare systems. DACBT
network algorithms may identify brain tumors early. classified brain tumors using brain magnetic
Segmentation improved brain tumor detection and resonance (MR) imaging data and an upgraded
categorization. The study trained models to reliably CNN. Adding data and transfer learning enhanced
identify brain cancers using MRI utilizing the model's categorization. The suggested model
convolutional neural networks (CNNs) and VGG-16 outperformed baseline brain cancer diagnostic
architecture. Popular for its simplicity and speed, methods in IoT-healthcare systems in experiments.
VGG-16 developed convolutional feature maps that This research improves brain tumor classification
were categorized to find tumor locations. The accuracy and efficiency using deep learning.
proposed method was tested on 253 MRI brain Dhakshnamurthy et al. (2024) classified
scans, 155 of which showed cancers. CNN's 96% brain cancers using transfer learning [13].
accuracy and 91.78% F1-score, VGG-16's 98.5% Understanding the time-consuming nature of brain
and 92.6%, and the ensemble model's 98.14% tumor identification and the limitations of
outperformed classical methods. The project closes conventional techniques in managing expanding
with deep learning brain tumor detection data volume, the scientists researched deep learning

6260
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

approaches to build automated systems for precise of tumors, categorization of tumors, segmentation of
and efficient diagnosis. The study found AlexNet, tumors, and prediction of survival rates. It is
VGG16, and ResNet-50 effective. The authors then necessary to collect data, perform preprocessing, and
created a hybrid VGG16–ResNet-50 model with make improvements to the Convolutional
99.98% accuracy, sensitivity, specificity, and F1 Normalized Mean Filter in order to implement the
score. The framework can accurately identify approach for the BraTS dataset. Multi-class
cerebral neoplasms when compared to other models, classification is accomplished by the utilization of
showing the promise of transfer learning in medical the cutting-edge DBT-CNN classifier model. A
imaging for brain tumor identification and survival rate prediction is made using a logistic
classification. Dhiman et al. (2022) proposed a regression model after the tumor has been delineated
machine learning-based hybrid CNN model for and features have been extracted using RU-Net2+.
medical image processing tumor detection [14]. The When it came to classification accuracy, tumor
study quickly extracted valuable malignant tumor segmentation precision, and survival rate prediction,
oncology medical data from electronic clinical the experimental findings were superior to the
medical records. The study provided a standard standards that are now in place. LGG tumors had a
extraction method for initial tumor site, size, and classification accuracy of 99.28%, whereas HGG
metastatic sites to address tumor-related medical malignancies had a classification accuracy of
events. Also, key-based and pseudo-data-generation 99.51%. The HGG had a tumor segmentation
algorithms improved transfer learning across tumor- accuracy of 98.39%, whereas the LGG had a 99.1%
related medical event extractions. Multiple accuracy. Long-term patient survival rates were
CCKS2020 dataset experiments demonstrated that predicted by the RU-Net2+ algorithm to be 85.71%,
the recommended approach placed third in clinical medium-term survival rates to be 72.72%, and short-
medical event extraction and assessment of the term survival rates to be 61.54%, with Mean Squared
electronic medical record. Errors of 0.13, 0.21, and 0.31 respectively. These
Deep learning and machine learning were findings provide medical practitioners with
utilized by Senan et al. in 2022 in order to investigate information that assists them in making decisions
the early identification of brain tumors using MRI regarding the treatment of brain tumors and
images [15]. As a result of the fact that cancer is one demonstrate that automated brain tumor
of the most aggressive and life-threatening diseases, identification may enhance patient care.
the purpose of the study was to improve patient Srinivasan et al. (2023) used deep learning
survival by properly detecting brain tumors. With to grade cancers from brain magnetic resonance
advances in deep and machine learning, computer- imaging [17]. In order to increase tumor
aided diagnostic systems can help doctors make identification accuracy, the study developed a
accurate diagnoses. The study integrated deep unique automated detection and classification
learning methods like AlexNet and ResNet-18 with approach. MRI scans were pre-processed,
classic machine learning methods like SVM for segmented, feature-extracted, and classified. During
brain tumor classification and diagnosis. Brain the pre-processing stage of magnetic resonance
tumor MRI pictures were improved using the imaging (MRI), an adaptive filter was utilized to
average filter. Deep convolutional layers extracted eliminate background noise. The local-binary grey
robust and crucial deep features, followed by level co-occurrence matrix (LBGLCM) was utilized
SoftMax and SVM classification. The 3,060 pictures for feature extraction, and enhanced fuzzy c-means
were sorted into three tumor kinds and one normal clustering was utilized for image segmentation. The
class. All systems performed well, but the images obtained from an MRI were categorized as
AlexNet+SVM hybrid approach had the highest either normal or glioma using a convolutional
accuracy (95.10%), sensitivity (95.25%), and recurrent neural network (CRNN). In terms of
specificity (98.50%). This study shows that hybrid accuracy, specificity, and sensitivity, the
methods can detect brain tumors early, advancing recommended method for classifying brain tumors
medical imaging and diagnostics. performed better than BP, U-Net, and ResNet. It
RU-Net2+ is a deep learning system that achieved 98.17 percent, 91.34%, and 98.79 percent
was created by Zaitoon and Syed (2023) for the respectively. In order to demonstrate that the CRNN
purpose of brain tumor segmentation and survival technique enhanced the accuracy of brain image
rate prediction [16]. The process of diagnosing and classification, the research utilized 620 testing and
treating brain tumors is challenging, but deep 2480 training MRI images taken from the
learning may make it possible to automate the REMBRANDT dataset that were employed.
process. The framework incorporates the detection

6261
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

ZainEldin et al. (2023) classified brain revolutionize medical imaging analysis [6–17].
cancers using deep learning and sine-cosine fitness However, despite these advancements, several gaps
gray wolf optimization [26]. Due to increased patient in the field remain to be addressed. One such
data volume, the study intended to develop reliable technical gap is the need for more robust and
and efficient automated brain tumor diagnosis generalizable deep learning models that can
methods, which are time-consuming and rely on effectively handle diverse patient populations and
radiologists. Pre-trained CNN models are used for imaging modalities. Additionally, there is a pressing
quicker brain tumor identification and segmentation need to develop more effective and secure federated
using deep learning. BCM-CNN used ADSCFGWO learning frameworks to address privacy concerns
to tune CNN hyperparameters. Hyperparameter associated with medical data sharing [7].
adjustment and model training with Inception- Furthermore, the establishment of standard
ResnetV2, a popular pre-trained model, provided evaluation protocols and benchmark datasets is
binary outcomes (0: Normal, 1: Tumor) to improve essential to enable fair comparisons between
brain tumor diagnosis. ADSCFGWO improved different automated brain tumor analysis systems
hyperparameters utilizing sine cosine and grey wolf [7]. Lastly, extensive validation studies are required
strengths. BCM-CNN obtained 99.98% accuracy on to evaluate the practicality and clinical utility of
the BRaTS 2021 Task 1 dataset, indicating these automated systems in real-world settings [8].
hyperparameter change increases CNN To overcome these technological shortcomings, our
performance. Butterfly optimization algorithm research provides an improved deep learning
(BOA) is a nature-inspired metaheuristic for global framework for brain tumor detection and
optimization presented by Arora and Singh (2019) classification using advanced methodologies. The
[28]. Computer scientists discovered more efficient approach uses Convolutional Neural Networks
ways to solve multidimensional and multimodal (CNNs) for precise tumor detection and hybrid
real-world problems. Metaheuristic algorithms architectures like U-Net to handle complicated
inspired by nature outperform traditional approaches sections and fine textures during down-sampling for
for comparable problems. BOA uses fragrance like robust segmentation [13]. Additionally, VGG-16 is
butterflies to discover nectar and mates. It was tested utilized for tumor grade classification, with relevant
on 30 benchmark functions and compared to other features extracted from segmented tumor images to
metaheuristic algorithms. BOA also addressed three achieve high accuracy [8]. To overcome complexity
old technical problems: spring, welded beam, and and overfitting issues, the Butterfly Optimization
gear train design. BOA was more efficient than algorithm is hybridized with VGG-16, enhancing the
previous metaheuristic algorithms, implying a global model's performance in classifying tumor grades
optimization revolution. Makhadmeh et al. (2023) [26]. By addressing these technical gaps, our
conducted a study of the butterfly optimization research contributes to the advancement of
algorithm (BOA), which is a metaheuristic swarm- automated brain tumor analysis systems, offering
based optimization approach that has proved to be improved accuracy and efficiency in tumor detection
successful in recent times [29]. Scholars have and classification. Moreover, the deployment of our
studied the BOA owing of its few adaptive model on JETSON Orin hardware ensures its
characteristics and good exploration-exploitation performance, portability, and energy efficiency,
balance. In a short time, BOA has been widely used further enhancing its practical applicability in
for optimization issues in several fields. The study clinical settings. Through multidisciplinary
begins by explaining BOA's fundamentals and collaboration and innovative research, we aim to
optimization ideas. It then examines BOA's translate these advancements into therapeutically
mathematical model and motivation, using an useful solutions while upholding patient privacy,
example to demonstrate its usefulness. Adaptation data security, and regulatory compliance.
forms divide the examined research into original,
modified, and hybridized. The primary BOA
applications and their pros and cons in optimization
situations are presented. The report finishes with a
summary and offers future research on BOA's uses
and developments.

Several deep learning research have

improved brain tumor identification and
classification, showing their potential to

6262
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

3. METHODOLOGY
3.1 Proposed Methodology

Figure 1: Proposed Work Flow of Brain Tumor Detection Model

Figure 1 shows the proposed Work Flow Br35H[28], figshare[30], and the SARTAJ[31]
of Brain Tumor Detection Model. The proposed dataset for brain tumor identification and
methodology integrates Convolutional Neural classification. This dataset is well prepared and
Networks (CNNs) with U-Net for segmentation annotated to allow thorough study and model
and VGG-16 architecture with Butterfly training; it contains 7023 MRI pictures of the
Optimization for feature extraction. Data human brain. This dataset offers a broad picture of
preprocessing ensures standardization and noise brain tumor forms and pathophysiology. It
removal in brain MRI images. CNN+U-Net comprises four separate classes: glioma,
facilitates precise tumor segmentation by meningioma, no tumor, and pituitary. In particular,
integrating low and high-level features. VGG-16 the no-tumor class guarantees a complete depiction
extracts intricate image features, optimized by of brain areas free of tumors by using pictures
Butterfly Optimization for reduced complexity. drawn entirely from the Br35H dataset. Figure 2
Classification distinguishes tumor and non-tumor shows the overview of Brain Tumor Data.
regions using extracted features. Evaluation Upon careful examination, it was
metrics and clinical validation ensure the efficacy observed that the glioma class images from the
and applicability of the proposed framework in SARTAJ dataset exhibited categorization
accurate brain tumor diagnosis and treatment inconsistencies, prompting their exclusion from
planning. the dataset. To address this issue, the glioma class
images were replaced with those sourced from the
3.1.1 Dataset Description figshare repository, ensuring the integrity and
Our research relies on a dataset[29] that is accuracy of class labels for model training and
produced from a mixture of three main sources: evaluation.

6263
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

Figure 2: Overview Of Brain Tumor Data

Table 1: Dataset Attributes For Brain Tumor Detection And Classification

Category No. Attribute Name
1 Image ID
2 Tumor Type
3 Patient ID
4 Image Dimensions
5 MRI Machine Vendor
6 MRI Machine Model
Basic Image Features
7 Image Modality
8 Image Orientation
9 Image Resolution
10 Image Acquisition Parameters
11 Image Contrast
12 Tumor Localization
1 Tumor Segmentation
2 Tumor Grade
3 Tumor Size
4 Tumor Morphology
5 Surrounding Tissue Characteristics
6 Tumor Enhancement
Content Related Features
7 Edema Presence
8 Necrosis Presence
9 Tumor Hemorrhage
10 Peritumoral Edema
11 Tumor Boundary Clarity
12 Tumor Growth Rate

6264
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

Table 1 presents the attributes encompassed where Width and Height denote the dimensions of
within the brain tumor. MRI dataset utilized for the image. This normalization technique ensures that
detection and classification tasks. The attributes are images are proportionally scaled, enhancing the
categorized into two main sections: Basic Image efficacy of subsequent feature extraction processes.
Features and Content Related Features. 3.1.4 Feature Extraction:
The distribution of images across classes is as Feature extraction is a crucial step in our
follows: methodology, aimed at capturing discriminative
➢ Glioma: 300 files spatial and spectral features from brain tumor
➢ Meningioma: 306 files images. Two main algorithms are employed for
➢ No tumor: 405 files feature extraction: the Multi-task Network and the
Global Net architecture.
➢ Pituitary: 300 files
3.2 Multi-task Network:
Additionally, a subset of 1,311 images from the
The Multi-task Network utilizes three
dataset is reserved for use in testing models. By
concurrent convolutional neural networks (CNNs) to
doing so, the model's performance may be validated
extract and process features with resilience. Table 2
using data that has not been seen before. The
presents a comprehensive summary of the design,
collection includes meticulously annotated images
specifying the quantity of filters, kernel sizes,
that indicate the presence or categorization of tumors
strides, and activation functions utilized in each
based on ground truth. This facilitates the
layer.
development of robust and precise deep-learning
𝑍 =𝑊∗𝑋+𝑏
models for the processing of medical images.
where Z represents the output feature map, W
Further, we manually categorized the data into the
denotes the convolutional filter weights, X denotes
following groups: tumor, tumor with grade II, tumor
the input image, and b represents the bias term. By
with grade III, tumor with grade IV, and no-tumor.
leveraging multiple CNNs in parallel, our model
Improved accuracy in tumor classification and
effectively captures diverse spatial and spectral
grading is a direct result of this manual separation,
features inherent in brain tumor images, enabling
which in turn allows for the development of more
comprehensive analysis and classification.
trustworthy deep-learning models for the detection
3.3 Global Net (U-Net) Architecture:
and classification of brain tumors.
The Global Net architecture, based on the
Our study results will be reliable and
renowned U-Net architecture, is specifically
informative because we used this vast dataset that
designed for semantic segmentation tasks. Table 3
was carefully picked. More than that, it enables us to
outlines the architectural details, including the
build therapeutically useful, high-powered deep-
number of filters, kernel sizes, strides, and activation
learning models for detecting and categorizing brain
functions employed in each layer. This architecture
cancers.
comprises both downsampling and upsampling
3.1.2 Feature Extraction & Preprocessing
layers, facilitating feature extraction and high-
In the early stages of our method, it is very
resolution feature map reconstruction.
important to carefully prepare the input data so that
By incorporating U-Net architecture into
our deep learning models can find and label brain
our model, we enhance its ability to capture intricate
tumors as well as possible. This section outlines the
structural details and spatial relationships within the
preprocessing techniques employed, followed by an
brain tumor images, thereby improving
in-depth discussion of the feature extraction process
segmentation accuracy and classification
using the Multi-task Network and the Global Net
performance.
architectures.
Together, the preprocessing techniques and
3.1.3 Preprocessing Techniques:
feature extraction algorithms outlined in this section
Before feature extraction, input MRI
form the foundation of our robust and efficient
images undergo preprocessing steps to standardize
approach to brain tumor detection and classification.
their aspect ratios and dimensions. Aspect ratio
By optimizing both data representation and feature
normalization and resizing techniques are applied to
extraction processes, we aim to develop a highly
ensure uniformity across the dataset, mitigating
accurate and clinically relevant deep learning model
potential distortions and facilitating consistent
capable of accurately identifying and classifying
model input. Mathematically, aspect ratio
brain tumors with unprecedented precision and
normalization is represented as:
𝑊𝑖𝑑𝑡ℎ reliability.
𝐴𝑠𝑝𝑒𝑐𝑡 𝑅𝑎𝑡𝑖𝑜 =
𝐻𝑒𝑖𝑔ℎ𝑡

6265
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

Table 2: Multi-Task Network and U-Net Architecture for Brain Tumor Segmentation

Layer Type Output Shape Parameters

InputLayer (None, 128, 128, 3) 0
Lambda (None, 128, 128, 3) 0
Conv2D (None, 128, 128, 8) 216
BatchNormalization (None, 128, 128, 8) 512
Activation (None, 128, 128, 8) 0
Conv2D (None, 128, 128, 8) 576
BatchNormalization (None, 128, 128, 8) 512
Activation (None, 128, 128, 8) 0
MaxPooling2D (None, 64, 64, 8) 0
Conv2D (None, 64, 64, 16) 1152
BatchNormalization (None, 64, 64, 16) 256
Activation (None, 64, 64, 16) 0
Conv2D (None, 64, 64, 16) 2304
BatchNormalization (None, 64, 64, 16) 256
Activation (None, 64, 64, 16) 0
MaxPooling2D (None, 32, 32, 16) 0
Conv2D (None, 32, 32, 32) 4608
BatchNormalization (None, 32, 32, 32) 128
Activation (None, 32, 32, 32) 0
Conv2D (None, 32, 32, 32) 9216
BatchNormalization (None, 32, 32, 32) 128
Activation (None, 32, 32, 32) 0
MaxPooling2D (None, 16, 16, 32) 0
Conv2D (None, 16, 16, 64) 18496
Activation (None, 16, 16, 64) 0
Conv2D (None, 16, 16, 64) 36928
BatchNormalization (None, 16, 16, 64) 256
Activation (None, 16, 16, 64) 0
MaxPooling2D (None, 8, 8, 64) 0
Conv2D (None, 8, 8, 128) 73856
Activation (None, 8, 8, 128) 0
Conv2D (None, 8, 8, 128) 147584
BatchNormalization (None, 8, 8, 128) 512
Activation (None, 8, 8, 128) 0

6266
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

MaxPooling2D (None, 4, 4, 128) 0

Conv2D (None, 4, 4, 256) 295168
Activation (None, 4, 4, 256) 0
Conv2D (None, 4, 4, 256) 590080
BatchNormalization (None, 4, 4, 256) 1024
Activation (None, 4, 4, 256) 0
MaxPooling2D (None, 2, 2, 256) 0
Conv2D (None, 2, 2, 512) 1180160
Activation (None, 2, 2, 512) 0
Conv2D (None, 2, 2, 512) 2359296
BatchNormalization (None, 2, 2, 512) 2048
Activation (None, 2, 2, 512) 0
MaxPooling2D (None, 1, 1, 512) 0
Conv2D (None, 1, 1, 1024) 4719616
Activation (None, 1, 1, 1024) 0
Conv2D (None, 1, 1, 1024) 9438208
BatchNormalization (None, 1, 1, 1024) 4096
Activation (None, 1, 1, 1024) 0
Conv2DTranspose (None, 2, 2, 512) 2097664
Concatenate (None, 2, 2, 1024) 0
Conv2D (None, 2, 2, 512) 4719104
Activation (None, 2, 2, 512) 0
Conv2D (None, 2, 2, 512) 2359296
BatchNormalization (None, 2, 2, 512) 2048
Activation (None, 2, 2, 512) 0
BatchNormalization (None, 2, 2, 512) 2048
Dense (None, 2, 2, 128) 65664
Dropout (None, 2, 2, 128) 0
Conv2DTranspose (None, 4, 4, 256) 131

Table 2 presents the Multi-Task Network In order to find the best answers, it mimics the way
and U-Net Architecture for Brain Tumor butterflies naturally move through their
Segmentation. This combined table represents the surroundings. By mimicking the dynamics of
multi-task network architecture and the U-Net butterfly movements, this algorithm offers a
architecture used in the proposed system for brain powerful approach to solving complex optimization
tumor segmentation. problems, including brain tumor detection and
classification in medical imaging analysis. At the
3.4 Butterfly Optimization Algorithm start of the optimization process, the search area is
The Butterfly Optimization algorithm is a filled with a random group of butterflies. Each
new type of metaheuristic optimization method that butterfly is a possible answer to the optimization
was inspired by the way butterflies flap their wings. problem. The positions of these butterflies

6267
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

correspond to candidate solutions, and their algorithm incorporates randomness to facilitate

movement within the search space is governed by exploration, ensuring that the search process remains
mathematical equations derived from the principles diverse and avoids premature convergence to
of butterfly behavior. Central to the Butterfly suboptimal solutions. Let's denote the position of the
Optimization algorithm is the evaluation of an i-th butterfly as 𝑥 , the best position found so far as
objective function that quantifies the fitness or 𝑥𝑏𝑒𝑠𝑡, and the attraction coefficient as 𝛽. The update
suitability of each potential solution. Most of the rule for the Butterfly Optimization algorithm can be
time, this target function is set based on the needs of represented as:
the optimization situation. When brain tumors are 𝑥 (𝑡 + 1) = 𝑥 (𝑡) + 𝛽 ∗ (𝑥𝑏𝑒𝑠𝑡(𝑡) − 𝑥 (𝑡)) + 𝛼
being found and categorized, the objective function ∗ (𝑥 (𝑡) − 𝑥 (𝑡))
includes performance measures like classification where 𝛼 is the step size, and 𝑥 (𝑡) is the position of
accuracy, sensitivity, specificity, or other measures a randomly selected butterfly at time 𝑡.
that are relevant to the job at hand. The derivative of the loss function with respect to the
Figure 3 shows the Workflow Design of attraction coefficient β can be represented as:
Butterfly Optimization Algorithm. The core of the 𝜕𝐿/𝜕𝛽 = 𝜕𝐿/𝜕𝑥 ∗ 𝜕𝑥 /𝜕𝛽
Butterfly Optimization algorithm lies in the iterative = 𝜕𝐿/𝜕𝑥 ∗ (𝑥𝑏𝑒𝑠𝑡 − 𝑥 )
process of updating the positions of butterflies based where 𝜕𝐿/𝜕𝑥 is the derivative of the loss function
on their current fitness evaluations. This movement with respect to the position of the 𝑖-th butterfly, and
is guided by a combination of exploitation and 𝑥𝑏𝑒𝑠𝑡 − 𝑥 is the difference between the best
exploration strategies, wherein butterflies balance position found so far and the current position of the
the exploitation of promising regions of the search
𝑖-th butterfly.
space with the exploration of new areas. The
facilitating the discovery of novel and potentially
superior solutions. This dual-mode search strategy
enhances the algorithm's robustness and scalability
across diverse optimization landscapes. The
positions of butterflies are iteratively updated based
on the fitness evaluations obtained from the
objective function. This updating process involves a
combination of deterministic and stochastic
components, wherein butterflies adjust their
positions to converge towards regions of higher
fitness. The magnitude and direction of these
adjustments are determined by mathematical
equations that encapsulate the dynamics of butterfly
movement and exploration.
The movement of butterflies within the
search space is governed by the following equations:

𝑥 (𝑡 + 1) = 𝑥 (𝑡) + 𝑣 (𝑡)
𝑣 (𝑡 + 1) = 𝑣 (𝑡) + 𝑐 . 𝑟 . (𝑥 ∗ − 𝑥 (𝑡))
+ 𝑐 . 𝑟 . (𝑥 ∗∗ − 𝑥 (𝑡))
Where 𝑥 (𝑡) and 𝑣 (𝑡) denote the position and
Figure 3: Workflow Design of Butterfly Optimization velocity of the 𝑖 − 𝑡ℎ butterfly along the 𝑑-th
Algorithm dimension at time 𝑡,respectively.𝑥 ∗ and 𝑥 ∗∗
represent the local and global best positions
A distinguishing feature of the Butterfly encountered by the butterfly,𝑐 and 𝑐 are
Optimization algorithm is its ability to perform both acceleration coefficients, and 𝑟 and 𝑟 are random
local and global search. Local search involves numbers sampled from uniform distributions. The
refining solutions in the vicinity of promising positions of butterflies are updated iteratively based
regions, leveraging the gradient information on their velocities, ensuring convergence towards
obtained from the objective function evaluations. regions of higher fitness. The position update
Meanwhile, global search enables butterflies to equation is given by:
explore distant regions of the search space, thereby
𝑥 (𝑡 + 1) = 𝑥 (𝑡) + 𝑣 (𝑡)

6268
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

Algorithm 1

Step 1: Generate random positions for each butterfly within the search space.
Step 2: Evaluate the objective function for each butterfly's position.
Step 3: Update the position and velocity of each butterfly using:
Butterfly movement equation:𝑥 (𝑡 + 1) = 𝑥 (𝑡) + 𝑣 (𝑡)
Velocity update equation:
𝑣 (𝑡 + 1) = 𝑣 (𝑡) + 𝑐 . 𝑟 . (𝑥 ∗ − 𝑥 (𝑡)) + 𝑐 . 𝑟 . (𝑥 ∗∗ − 𝑥 (𝑡))
𝑐 and 𝑐 are acceleration coefficients, 𝑟 and 𝑟 are random numbers.
Step 4: Perform local search around the best positions.
Step 5: Explore new regions in the search space.
Step 6: Update the positions of butterflies based on their velocities.
Termination:
Step 7: Repeat steps 2-5 until the termination criterion is met.

The above algorithm 1 is the Butterfly Our initial step involves adapting the
Optimization Algorithm (BOA). It initializes a conventional VGG-16 architecture to accommodate
swarm of butterflies in a search space and evaluates reduced weights obtained through Butterfly
their positions against an objective function. Optimization. Mathematically, this modification can
Through iterative updates guided by acceleration be expressed as:
coefficients and random numbers, butterflies 3.7 Weight Adjustment
navigate toward promising solutions while exploring 𝑊 =𝑊 ×𝛼
new regions. This process involves both local Where 𝑊 represents teh original weights of
refinement and global exploration to optimize the VGG-16, and 𝛼 denotes the weight reduction factor
objective function. Positions are continuously obtained through Butterfly Optimization.
updated based on butterfly velocities until a Using the Butterfly Optimization
termination criterion is met, leading to efficient technique, the VGG-16 model's weights are
optimization of complex problems. optimized iteratively. To fine-tune the model's
parameters for the goal of brain tumor detection and
3.5 Implementing Hybrid Visual Geometry classification, this optimization procedure involves
Group-16 with Reduced Weights via Butterfly modifying the weights depending on their
Optimization contribution to the overall classification accuracy.
In our pursuit of refining the efficacy of Our goal is to improve the model's diagnosis
brain tumor detection and classification, we accuracy by repeatedly tweaking the weights until it
implement a novel approach leveraging the synergy can better distinguish between tumor and non-tumor
between the renowned Visual Geometry Group-16 areas.
(VGG-16) architecture and the Butterfly Mathematically, the weight adjustment can be
Optimization algorithm. This fusion aims to enhance formulated as:
the discriminative power of VGG-16 while 𝑊 = 𝑊 + 𝛥𝑊
mitigating the computational burden associated with Where 𝑊 represents the current weights and 𝛥𝑊
its standard implementation. denotes the change in weights computed using
3.6 Architecture Modification of VGG-16 Butterfly Optimization.
The foundational step involves adapting the
conventional VGG-16 architecture to accommodate
reduced weights obtained through the Butterfly
Optimization algorithm (algorithm 2). This
modification streamlines the model's complexity
while preserving its ability to extract relevant
features from medical imaging data. By
incorporating reduced weights, we aim to achieve a
balance between model performance and
computational efficiency, ensuring optimal
utilization of computational resources.

6269
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

Algorithm 2

Step 1: Initialize the VGG-16 architecture with random weights.

Step 2: Generate a swarm of butterflies within the search space.
Step 3: Perform Butterfly Optimization to obtain a reduction factor 𝛼.
Update the weights of VGG-16 using the reduction factor:
For each layer 𝑙 in VGG-16:
Obtain the original weight matrix 𝑊 .
Multiply 𝑊 by to obtain the reduced weight matrix 𝑊 .
Step 4: Feed the modified VGG-16 architecture with the reduced weights using labeled brain tumor
datasets.
Utilize standard optimization techniques of stochastic gradient descent to update the weights and
biases of the network.
Step 5: Iterate over the training dataset multiple times to minimize the classification loss.

Figure 4: Sequence Flow For The Proposed Hybrid Visual Geometry Group-16 With Reduced Weights Via Butterfly
Optimization

The figure 4 illustrates the sequence flow and classification, this figure offers a structured
for the proposed Hybrid Visual Geometry Group-16 visual representation of the process.
(VGG-16) with Reduced Weights via Butterfly 4. EXPERIMENTATION RESULTS AND
Optimization. It outlines the sequential interactions DISCUSSION
within the algorithm, commencing with the Experiments were conducted on a dataset
initialization phase, where the VGG-16 architecture consisting of 7023 brain MRI images to see how
is initialized, and Butterfly Optimization is initiated. well the suggested hybrid method worked.
The reduction factor obtained from Butterfly
Optimization is then applied to update the weights of 4.1 Experimental Setup
VGG-16. The design then goes on to classify and A computer with an Intel Core i7 processor,
extract features, and finally, during training, it 16 GB of RAM, and an NVIDIA GeForce GTX 1080
updates the network's biases and weights using Ti graphics card was used for the experiments, as
labeled datasets of brain tumors. Lastly, validation was an NVIDIA Jetson AGX Orin module, a low-
datasets are used to evaluate the model's power AI computing platform designed for edge AI
performance in terms of classification accuracy. In applications. The proposed approach was
order to better understand how to apply the implemented using the Keras deep learning
suggested hybrid method for brain tumor detection framework with a TensorFlow backend. The
approach's adaptability and potential for deployment

6270
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

in multiple situations were demonstrated by our platform. Figure 5 shows the Hardware
evaluation of its performance on both a high- Specification Used for Experimentation.
performance computer and a low-power edge AI

Figure 5: Hardware Specification Used For Experimentation

4.2 Dataset 60:20:20. The dataset was annotated manually, with

The dataset used in the experiments each image labeled as either tumor, tumor grade II,
consisted of 7023 brain MRI images, divided into tumor grade III, tumor grade IV, or no tumor. Figure
training, validation, and testing sets in a ratio of 6 shows the Annotated and Processed Dataset.

Figure 6 Annotated and Processed Dataset

4.3 Experimental Design 4.3.1 Hyperparameter Tuning

The studies were planned to assess the A grid search technique was used to
accuracy, sensitivity, specificity, Dice coefficient, optimize the hyperparameters of the suggested
and Jaccard coefficient of the proposed hybrid approach. Learning rate, batch size, and epoch count
method. The approach was compared with existing were among the hyperparameters fine-tuned. The
approaches, including the standard VGG-16 method's efficacy on the validation set informed the
architecture and other deep learning models. selection of the best hyperparameters.

6271
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

Figure 7: Hyperparameter Tuning Results and Learning Rate

The outcomes of tuning the shows the Optimized Parameters for the Proposed
hyperparameters of a deep learning model through Model.
the utilization of various combinations of dropout Table 3: Optimized Parameters for the Proposed Model
rates, learning rates, loss functions, and optimizers Optimized
Hyperparameter
are illustrated in Figure 7. The plot displays the Value
relationships between these hyperparameters and the Learning Rate 0.00123
resulting model performance, measured by the loss Batch Size 32
function. The color bar indicates the loss value, with Number of Epochs 150
lower values indicating better model performance. Dropout Rate 0.25
The scatter points represent individual Weight Decay 0.0001
hyperparameter combinations, with the size of each Momentum 0.9
point corresponding to the dropout rate. The y-axis Activation Function (Hidden
ReLU
denotes the optimizer executed, while the x-axis Layers)
represents the learning rate. The plot can be used to Activation Function (Output
Sigmoid
Layer)
identify the most promising hyperparameter
Number of Hidden Layers 3
combinations and to visualize the relationships
between the different hyperparameters. Table 3 Number of Neurons (Hidden
128, 64, 32
Layers)

Figure 8 Accuracy Comparison of Deep Learning Models with and without Butterfly Optimization on Local PC and
Nvidia Jetson AGX ORIN

6272
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

Figure 8 shows the accuracy Comparison of and Proposed VGG-16. In this graph, we can
Deep Learning Models with and without Butterfly observe how well each model performed with and
Optimization on Local PC and Nvidia Jetson AGX without Butterfly Optimization, a unique
ORIN. On two separate hardware platforms, a local optimization method developed to boost the
PC (NVIDIA GTX 1080) and an Nvidia Jetson AGX efficiency of DL models. The outcomes show that
ORIN, this graphic compares the accuracy of Butterfly Optimization is a powerful tool for
multiple deep learning models, such as Simple CNN, enhancing the precision of deep learning models on
Mix-pooling CNN with FCRF Segmentation, Fine- both hardware types.
tuned YOLOv7, BCM-CNN, Inception+resnetV2,

Figure 9 Performance Metrics of Deep Learning Models

Figure 9 shows the Performance Metrics of whose performance metrics are compared. Among
Deep Learning Models. Simple CNN, Mix-pooling the measures are F1 Score, Accuracy (ACC),
CNN with FCRF Segmentation, Fine-tuned Precision, and Recall. The outcomes prove that
YOLOv7, BCM-CNN, Inception+resnetV2, and Butterfly Optimization is a powerful tool for
Proposed VGG-16 (with and without Butterfly enhancing the efficiency of deep learning models.
Optimization) are some of the deep learning models

6273
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

Figure 10 Performance Comparison of Segmentation Algorithms

This figure 10 compares the performance of (IOU), Dice Coefficient, and Jaccard Distance. The
various segmentation algorithms, including U-Net, findings show that the CNN+UNet model works
FCN, SegNet, CNN+UNet (Proposed), well, with competitive performance on Dice
CNN+ResNet50, DeepLab, and PSPNet. The Coefficient and Jaccard Distance and excellent
metrics include Accuracy, Intersection over Union accuracy and IOU.

6274
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

Figure 11: The Output Predicted by Our Proposed System

Figure 11 illustrates the output predicted by tumor regions with high fidelity. Furthermore, the
our proposed system for brain tumor detection and classification of tumor grades using VGG-16,
classification, run on a local PC. The image enhanced by the Butterfly Optimization algorithm,
showcases the segmented tumor regions accurately ensures accurate characterization of tumor severity.
identified by the optimized deep learning The integration of these methodologies results in a
framework. Utilizing advanced techniques such as comprehensive and efficient system for brain tumor
Convolutional Neural Networks (CNNs) and hybrid analysis, demonstrating the feasibility of deploying
architectures like U-Net, the system achieves precise deep learning models on local hardware for medical
segmentation, distinguishing tumor areas from non- imaging diagnostics.

6275
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

Figure 12 Predicted Outputs on NVIDIA Jetson AGX Orin Module

Figure 12 presents the predicted outputs of tumor segmentation and grade classification. The
our brain tumor detection and classification system predicted outputs include critical, mild, and
deployed on the NVIDIA Jetson AGX Orin module. moderate-grade tumors, with corresponding sizes of
The image illustrates accurately segmented tumor 18360, 2688, and 4256, respectively. Additionally,
regions, showcasing the system's robust the absence of tumors is accurately identified,
performance in real-world hardware settings. ensuring reliable diagnosis and efficient deployment
Utilizing advanced deep learning techniques and in clinical environments. Table 4 shows the
hybrid architectures like U-Net and VGG-16 with Segmentation Models with Data Augmentation and
Butterfly Optimization, the system achieves precise Computational Efficiency.

Table 4: Segmentation Models with Data Augmentation and Computational Efficiency

Model Number of Training Inference Memory Data Augmentation Dataset

Parameters Time Time Usage

CNN+UNet 10M 2 hours 10ms 4GB Rotation (30°), Width Segmentation

(Proposed) Shift (0.1), Height Dataset
Shift (0.1), Shear
(0.2), Zoom (0.2)
CNN+ResNet50 20M 4 hours 20ms 8GB Rotation (30°), Width Segmentation
Shift (0.1), Height Dataset
Shift (0.1), Shear
(0.2), Zoom (0.2)

6276
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

Good convergence and generalization

performance are indicated by the suggested model's
training loss of 0.0374 and validation loss of 0.0377.
By reducing the loss function and obtaining correct
segmentation results, the loss curves show that the
suggested model is successful. The findings indicate
that the suggested method is capable of achieving
real-time performance on the NVIDIA Jetson AGX
Orin hardware, with an inference time of 300
milliseconds and a power consumption of 15 watts.
Table 5 shows the Hardware Performance Results.

Table 5 Hardware Performance Results

Power
Figure 13 Training and Validation Loss Curves for the Inference Consumption
Proposed Segmentation CNN+UNet Model Hardware Time (ms) (W)

Figure 13 shows the Training and Computer 150 250

Validation Loss Curves for the Proposed
Segmentation CNN+UNet Model. Here we can NVIDIA
Jetson AGX 300 15
observe the segmentation CNN+UNet model's loss
Orin
curves, both during training and validation. The blue
and orange loss curves, representing training and
validation, are shown against the epoch count.
5. CONCLUSION
The proposed optimized deep learning
framework for brain tumor detection and
classification represents a significant advancement
in medical image analysis, seamlessly integrating
Convolutional Neural Networks (CNNs), U-Net
architecture, and the Butterfly Optimization
algorithm. Through meticulous data preprocessing,
including aspect ratio normalization and resizing, the
framework ensures standardized input data, resulting
in a segmentation accuracy of 98% as measured by
the Dice coefficient and a classification accuracy of
73% with VGG-16. The hybridization with the
Butterfly Optimization algorithm further enhances
performance, achieving an overall effectiveness of
99.99%. Demonstrating resilience to class
imbalances and complexity, the framework shows
suitability for real-world applications in clinical
Figure 14: ROC Curve Values for the Proposed Hybrid
VGG-16 Model with Reduced Weights via Butterfly
settings, facilitated by its implementation on
Optimization JETSON Orin hardware. Future research should
focus on refining the framework's architecture and
The suggested Hybrid VGG-16 model with algorithms, exploring novel techniques such as
weight reductions using Butterfly Optimization is ensemble learning and transfer learning, integrating
shown in Figure 14 along with the values of the ROC clinical expertise, and enhancing interpretability for
curve. The actual values of the thresholds, as well as broader adoption and impact. In conclusion, the
the true positive rate (TPR), false negative rate proposed framework offers a robust and efficient
(FNR), and true negative rate (FPR), are detailed in solution for brain tumor detection and classification,
the table. By plotting the ROC curve with these poised to revolutionize diagnostic practices and
numbers, we can see how well the suggested model improve patient outcomes in clinical settings.
works.

6277
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

REFERENCES [9]. 9. M. Kadkhodaei et al., "Automatic

[1]. K. Le Dinh Viet, K. Le Ha, T. N. Quoc, and V. segmentation of multimodal brain tumor
T. Hoang, "MRI Brain Tumor Classification images based on classification of super-
based on Federated Deep Learning," 2023 voxels," 2016 38th Annual International
Zooming Innovation in Consumer Conference of the IEEE Engineering in
Technologies Conference (ZINC), Novi Sad, Medicine and Biology Society (EMBC),
Serbia, 2023, pp. 131-135. Orlando, FL, USA, 2016, pp. 5945-5948.
[2]. A. Hasan, R. Toufiq and M. Z. Islam, [10]. S. A. Nagtode, B. B. Potdukhe, and P. Morey,
"Automated Multi-Class Brain Tumor "Two-dimensional discrete Wavelet transform
Classification for MRI Image Using Transfer and Probabilistic neural network used for brain
Learning," 2023 26th International Conference tumor detection and classification," 2016 Fifth
on Computer and Information Technology International Conference on Eco-friendly
(ICCIT), Cox's Bazar, Bangladesh, 2023, pp. Computing and Communication Systems
1-6. (ICECCS), Bhopal, India, 2016, pp. 20-26.
[3]. I. Ramírez, A. Martín and E. Schiavi, [11]. M. Nazir, M. A. Khan, T. Saba, and A.
"Optimization of a variational model using Rehman, "Brain Tumor Detection from MRI
deep learning: An application to brain tumor images using Multi-level Wavelets," 2019
segmentation," 2018 IEEE 15th International International Conference on Computer and
Symposium on Biomedical Imaging (ISBI Information Sciences (ICCIS), Sakaka, Saudi
2018), Washington, DC, USA, 2018, pp. 631- Arabia, 2019, pp. 1-5.
634. [12]. A. Banerjee, K. Jaiswal, T. Biswas, V. Sharma,
[4]. A. Gamal, K. Bedda, N. Ashraf, S. Ayman, M. M. Bal and S. Mishra, "Brain Tumor Detection
Abdallah and M. A. Rushdi, "Brain Tumor and Classification Using a Hyperparameter
Segmentation using 3D U-Net with Tuned Convolutional Neural Network," 2023
Hyperparameter Optimization," 2021 3rd 6th International Conference on Contemporary
Novel Intelligent and Leading Emerging Computing and Informatics (IC3I), Gautam
Sciences Conference (NILES), Giza, Egypt, Buddha Nagar, India, 2023, pp. 502-506.
2021, pp. 269-272. [13]. U. M and M. Baskar, "Machine Learning for
[5]. K. M. M. Dolaat, A. Erbad, and M. Ibrar, Brain Tumour Detection: A Systematic
"Enhancing Global Model Accuracy: Review of Segmentation and Classification
Federated Learning for Imbalanced Medical Methods," 2023 Intelligent Computing and
Image Datasets," 2023 International Control for Engineering and Business Systems
Symposium on Networks, Computers and (ICCEBS), Chennai, India, 2023, pp. 1-7.
Communications (ISNCC), Doha, Qatar, 2023, [14]. R. M. Prakash and R. S. S. Kumari,
pp. 1-4. "Classification of MR Brain Images for
[6]. N. M. Dipu, S. A. Shohan and K. M. A. Salam, Detection of Tumor with Transfer Learning
"Deep Learning Based Brain Tumor Detection from Pre-trained CNN Models," 2019
and Classification," 2021 International International Conference on Wireless
Conference on Intelligent Technologies Communications Signal Processing and
(CONIT), Hubli, India, 2021, pp. 1-6. Networking (WiSPNET), Chennai, India,
[7]. 7. T. P. Pries, R. Jahan, and P. Suman, "Review 2019, pp. 508-511.
of Brain Tumor Segmentation, Detection and [15]. F. P. Polly, S. K. Shil, M. A. Hossain, A.
Classification Algorithms in fMRI Images," Ayman and Y. M. Jang, "Detection and
2018 International Conference on classification of HGG and LGG brain tumor
Computational and Characterization using machine learning," 2018 International
Techniques in Engineering & Sciences Conference on Information Networking
(CCTES), Lucknow, India, 2018, pp. 300-303. (ICOIN), Chiang Mai, Thailand, 2018, pp.
[8]. 8. Y. Bhanothu, A. Kamalakannan and G. 813-817.
Rajamanickam, "Detection and Classification [16]. A. D, R. R, S. V, V. N and R. R, "Automated
of Brain Tumor in MRI Images using Deep Brain Tumor Classification and Detection
Convolutional Network," 2020 6th Using Modified Convolutional Neural
International Conference on Advanced Networks for Early Diagnosis," 2023
Computing and Communication Systems International Conference on Intelligent
(ICACCS), Coimbatore, India, 2020, pp. 248- Technologies for Sustainable Electric and
252.

6278
 Journal of Theoretical and Applied Information Technology
31st August 2024. Vol.102. No. 16
© Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

Communications Systems (iTech SECOM), [25]. Abdusalomov AB, Mukhiddinov M, Whangbo

Coimbatore, India, 2023, pp. 351-355. TK. Brain Tumor Detection Based on Deep
[17]. BeebiNaseeba, S. B. Nikhil, N. S. Nair, Learning Approaches and Magnetic
AsutoshDoppalapudi, N. P. Challa and R. Resonance Imaging. Cancers (Basel). 2023
Katta, "Classification and Detection of Brain Aug 18;15(16):4172. doi:
Tumors from Magnetic Resonance Imaging 10.3390/cancers15164172. PMID: 37627200;
Scans using Deep Transfer-Learning," 2023 PMCID: PMC10453020.
International Conference on Advances in [26]. ZainEldin H, Gamel SA, El-Kenawy E-SM,
Computing, Communication and Applied Alharbi AH, Khafaga DS, Ibrahim A, Talaat
Informatics (ACCAI), Chennai, India, 2023, FM. Brain Tumor Detection and Classification
pp. 1-7. Using Deep Learning and Sine-Cosine Fitness
[18]. H. Hu, X. Li, W. Yao and Z. Yao, "Brain Grey Wolf Optimization. Bioengineering.
Tumor Diagnose Applying CNN through 2023; 10(1):18.
MRI," 2021 2nd International Conference on [27]. Azzahra, T. S., Jessica Jesslyn Cerelia, Farid
Artificial Intelligence and Computer Azhar Lutfi Nugraha, & Anindya Apriliyanti
Engineering (ICAICE), Hangzhou, China, Pravitasari. (2023). MRI-Based Brain Tumor
2021, pp. 430-434. Classification Using Inception Resnet V2.
[19]. T. A. Jemimma and Y. J. Vetharaj, "Watershed Enthusiastic : International Journal of Applied
Algorithm based DAPP features for Brain Statistics and Data Science, 3(2), 163–175.
Tumor Segmentation and Classification," 2018 https://doi.org/10.20885/enthusiastic.vol3.iss2
International Conference on Smart Systems .art4
and Inventive Technology (ICSSIT), [28]. Hamada, A. (2020). Br35H :: Brain Tumor
Tirunelveli, India, 2018, pp. 155-158. Detection 2020 [Dataset]. Retrieved from
[20]. G. N, V. Pushpalatha, R. C, S. L and S. S, [https://paperswithcode.com/dataset/br35h-
"Brain Tumor Detection and Classification brain-tumor-detection-2020]
Using Deep Learning," 2023 Winter Summit [29]. Jyotismita Chaki, Marcin Wozniak, July 26,
on Smart Computing and Networks 2023, "Brain Tumor MRI Dataset", IEEE
(WiSSCoN), Chennai, India, 2023, pp. 1-6. Dataport, doi:
[21]. G. N, V. Pushpalatha, R. C, S. L and S. S, https://dx.doi.org/10.21227/1jny-g144.
"Brain Tumor Detection and Classification [30]. 30. Cheng, Jun (2017). brain tumor dataset.
Using Deep Learning," 2023 Winter Summit figshare. Dataset.
on Smart Computing and Networks https://doi.org/10.6084/m9.figshare.1512427.
(WiSSCoN), Chennai, India, 2023, pp. 1-6. v5
[22]. S. A. Swapnil and V. S. Girish, "Image Mining [31]. Chakrabarty, N., Kanchan, S., Bhuvaji, S.,
Methodology for Detection of Brain Tumor: A Kadam, A., Bhumkar, P., & Dedgev, S. (Year).
Review," 2020 Fourth International Brain Tumor Classification (MRI) Definition
Conference on Computing Methodologies and [Dataset]. Retrieved from
Communication (ICCMC), Erode, India, 2020, [https://www.kaggle.com/datasets/sartajbhuva
pp. 232-237. ji/brain-tumor-classification-mri]
[23]. M. Gurbină, M. Lascu and D. Lascu, "Tumor
Detection and Classification of MRI Brain
Image using Different Wavelet Transforms
and Support Vector Machines," 2019 42nd
International Conference on
Telecommunications and Signal Processing
(TSP), Budapest, Hungary, 2019, pp. 505-508.
[24]. Jie C., Luming Z., Naijie G., Xiaoci Z.,
Minquan Y., Rongzhang Y., Meng Q. A mix-
pooling CNN architecture with FCRF for brain
tumor segmentation. J. Vis. Commun. Image
Represent. 2019;58:316–322.

6279