International Journal of Computing and Digital Systems

ISSN (2210-142X)
Int. J. Com. Dig. Sys. 15, No.1 (Jan-24)
http://dx.doi.org/10.12785/ijcds/150106

A Stable Method For Brain Tumor Prediction In Magnetic

Resonance Images Using Fine-tuned XceptionNet
Shanmuga Sundari.M1 , Yeluri Divya2 , KBKS Durga3 , Vidyullatha Sukhavasi4 , M.Dyva Sugnana Rao5
and M.Sudha Rani6
1,2,3,4,5,6
Department of CSE, BVRIT HYDERABAD College of Engineering, Hyderabad, India

Received 4 May. 2023, Revised 20 Oct. 2023, Accepted 16 Nov. 2023, Published 1 Jan. 2024

Abstract: Brain tumors can be a life-threatening condition, and early detection is crucial for effective treatment. Magnetic resonance
imaging (MRI) is a valuable appliance for identifying the tumor’s location, but manual detection is a time-engrossing and flaws-prone
process. To overcome these challenges, computer-assisted approaches have been developed, and deep learning (DL) archetypes are now
being pre-owned in medical imaging to discover brain tumors maneuver MRI carbon copies. In this, we propose a deep convolutional
neural network (CNN) Xception net model for the efficient classification and detection of brain tumor images. We utilized the ”Br35H
:: Brain Tumor Detection 2020” dataset sourced from Kaggle, which encompasses 3000 MRI images of brain tumors, each with a file
size of 88 megabytes. The Xception net is a powerful CNN model that has shown promising results in various systems perceiving
exercise, in conjunction with medical illustration scrutiny. We fine-tuned the Xception net model using a dataset of Magnetic Resonance
Imaging (MRI) images of the brain, which were pre-processed and labeled by medical experts. To reckon the performance of our
prototype, we counselled dossier using a variety of interpretation criterion, including accuracy, precision, recall, and F1 score. Our
customs view that the urged model achieved high accuracy in classifying brain tumor images. The archetype’s strength to accurately and
efficiently classify and detect brain tumors using MRI images can significantly improve patient outcomes by enabling early detection
and treatment. Overall, our study demonstrates the persuasiveness of using the Xception net flawless for brain tumor ferreting out and
alloting using MRI images with 94% of accuracy performance. The proposed model has the potential to revolutionize the department
of salutary exemplify and improve patient outcomes for brain tumor treatment.

Keywords: Brain Tumor, Deep Convolution Neural Networks, Magnetic Resonance Imaging, XceptionNet

1. INTRODUCTION The exact causes of brain tumors are still unknown,

A brain tumor [1] refers to an heteromorphic sprouting but there are some risk factors that have been identified.
of cells in the brain or the surrounding tissues. It can be Exposure to radiation is one of the most significant risk
benign or malignant, and the latter is the more dangerous of factors, and it can increase the risk of developing a brain
the two. Brain tumors can develop in anyone, but they are tumor by up to 30%. Other risk factors include a family
more common in older adults. a variety of factors, including history of brain tumors, certain genetic disorders, and
the precise position and extent of the tumour, the symp- exposure to certain chemicals.
toms and indications of a brain tumour can change. Some
common symptoms include headaches, seizures, difficulty Diagnosing a brain tumor typically involves a combina-
speaking or moving, and changes in vision or hearing. tion of pictorial evaluations, such as CT scans, MRI scans,
and PET scans, as well as a biopsy, which involves remov-
There are several types of brain tumors, and they are ing a small sample of tissue from the tumor for analysis.
classified based on the type of cell they originate from. Once a brain tumor has been diagnosed, treatment options
For example, a glioma is a type of tumor that originates may include surgery [2], radiation therapy, chemotherapy, or
from glial cells, which are the cells that support and a combination of these approaches. The choice of treatment
nourish neurons in the brain. Another type of tumor is a will rely on the category of carcinoma, its location, and its
meningioma, which develops in the membranes that hem dimensions.
in the brain and spinal cord. Metastatic cysts are those that
have profusion from further parts of the anatomy to the One of the challenges in treating brain tumors is the lo-
brain. cation of the tumor itself. The brain is a delicate organ, and
surgery to remove a brain tumor can be risky, particularly if

E-mail address: sundari.m@bvrithyderabad.edu.in, divya.y@bvrithyderabad.edu.in, https:// journal.uob.edu.bh/

durgakbks@bvrithyderabad.edu.in, vidyullatha.1988@gmail.com, sugnanarao.m@bvrithyderabad.edu.in,
sudharani.m@bvrithyderabad.edu.in
68 Shanmuga Sundari M, et al.: Brain Tumor Prediction using Fine-tuned XceptionNet

Treatment can be difficult and can cause significant side

effects, and the uncertainty of the future can be overwhelm-
ing. However, there are resources available to help support
people with brain tumors and their families. During this
trying period, peer support networks, guidance, and other
resources can offer both emotional and practical help.
In general, brain tumors are a serious and often life-
threatening condition. Although the exact causes are still
unknown, researchers are making progress in developing
new treatments and improving existing ones. Living with a
brain tumor can be challenging, but with the right support
and care, people with brain tumors can live full and mean-
ingful lives. If you are experiencing symptoms that could
be related to a brain tumor, it is important to seek medical
attention right away. Initial identification and therapy can
enhance results and raise the likelihood of a full recovery.
Depending on where it is located and how big it is, a
brain tumour is a evolvement of abnormal sections in the
brain that can result in a variety of symptoms. It may be
benign or malignant, the latter of which is more hazardous
and possibly fatal.
Brain tumours are a common interest for investigation
Figure 1. Presence of tumour in brain in the handle of machine learning as a difficult diagnostic
problem. To reliably identify and categorise tumours in
pharmaceutical exemplifying, such as MRI or CT scans,
the tumor is located in a critical area of the brain. Radiation numerous algorithms and procedures have been developed.
therapy can be effective in killing cancer cells, but it can
also damage healthy brain tissue. Chemotherapy is another Deep learning is one such method, which entails teach-
option, but it can have significant side effects. ing neural networks to recognise tumor-related patterns
in medical pictures [4]. This method has demonstrated
In recent years, researchers have been exploring new promising outcomes in the very accurate detection of brain
treatments for brain tumors, including immunotherapy and tumours and has the potential to enhance diagnosis.
targeted therapy. Immunotherapy involves using the body’s
own immune system to fight the tumor, while shepherd However, there are significant ethical issues raised by
remedial treatment incorporates using narcotics that and the creation and use of these machine learning algorithms.
care, people with brain tumors can live full and meaningful For instance, how is data gathered and used? Who has
lives. If you are experiencing symptoms that could be access to the data needed to train these algorithms? How do
related to a brain tumor, it is important to seek medical [3] we make sure that these algorithms are applied responsibly,
attention right away. Initial identification and therapy can openly, and fairly, and that they don’t reinforce prejudice
enhance results and raise the likelihood of a full recovery. or injustice that already exists?
Depending on where it is located and how big it is, a Together, researchers and decision-makers must define
brain tumour is a evolvement of abnormal sections in the standards and best practises for the creation and application
brain that can result in a variety of symptoms. It may be of machine learning algorithms in healthcare in order to
benign or malignant, the latter of which is more hazardous address these problems. This entails ensuring that data is
and possibly fatal. gathered in an ethical manner, that algorithms are clear
and understandable, and that their application prioritises the
Brain tumours are a common interest for investigation safety and wellbeing of patients.
in the handle of machine learning as a difficult diagnostic
problem. To reliably identify and categorise tumours in Overall, the unearthing and summary of brain tumours
significantly pin point the carcinoma cells. These new proving expert systems has considerable promise for bet-
treatments are still in the early stages of development, but tering patient outcomes. To make sure that this technology
they hold promise for the future. is used in a responsible and advantageous way, it is crucial
that we approach it with caution and thought for the ethical
Living with a brain tumor can be challenging, both consequences.
for the person with the tumor and for their loved ones.

https:// journal.uob.edu.bh
 Int. J. Com. Dig. Sys. 15, No.1, 67-79 (Jan-24) 69

A mass or fleshing out of preternatural thaws in the brain Hardik J. Pandya discusses the creation of an automated
is avowed as a brain tumour. Either benign or malignant system that uses electrical resistivity measurements to dis-
can apply to it, with the latter being more aggressive and tinguish between healthy brain tissue and tumor-containing
cancerous. Deep learning algorithms have recently showed brain tissue. The difficulties in handling and characterising
promise in helping in the identification and diagnosis of priceless human brain biopsy tissues during surgery are
brain tumours. addressed by this approach. According to the study, there
are considerable differences between the electrical resistiv-
Neural networks are used in deep culturing, a sort of ity of healthy and cancerous brain regions, which raises
ML techniques, to learn from data. It has been used in many the possibility that it might be used as a third biomarker
different domains, including as natural language processing to distinguish between the two. The study also showed
and picture recognition. Deep learning algorithms can be that despite maintaining the resistivity trends seen in the
trained on medical imaging data to find patterns that might ordinary and tumour nodules, formalin mania increases the
point to the existence of a carcinoma in the case of brain electrical immunity of both ordinal and tumour cells [7].
tumours. According to scientists, the proposed automated system and
biochips may make it possible to learn more about the
The variety in how normal brain tissue appears is one heterogeneity of brain tissue and how it affects prognosis.
of the difficulties in identifying brain tumours. By learning
to distinguish between normal tissue and abnormal tissue Elisabeth Klint discusses the creation and assessment
based on patterns in the data, deep learning algorithms of a system that, during brain tumour surgery, integrates
[5] can assist in overcoming this difficulty. This can be practical dossier of PpIX-luminescence, microcirculation,
especially helpful for finding tiny tumours that can be and towel-slateness. The system consists of an optic exam-
challenging to find using conventional methods. ination, a ray Doppler system, a luminescence tool with a
CCD spectrometer for identifying PpIX peaks, and LabView
Deep learning can help with diagnosis and treatment software. Homonid dermis, brain tumour tissue, and immo-
planning in addition to detection. Deep learning algorithms bile fluorescing perceptible were used to gauge the system’s
can assist in identifying the kind, location, and size of efficiency. According to the findings, the system was able
the tumour by examining medical imaging data. Treatment to detect PpIX peaks in brain tumour tissue, which reduced
choices, such as whether surgery or radiation therapy is over time as a result of print bleaching [8]. Additionally,
required, can be guided by this information. the system’s electrical and ray safety for clinical application
The appositeness of deep learning to the identification was assessed. The method will then be put to the test during
and diagnosis of brain tumours is still in its infancy, despite necropsies and resections of clinical brain tumours.
its potential. The creation of robust algorithms that can Naveed Ilyas explains the most prevalent and quickly
generalise to new scenarios and the requirement for vast spreading types of brain tumours, gliomas, must be iden-
quantities of high-quality data are only two of the numerous tified and treated as soon as possible in order to in-
obstacles that must be addressed. Deep learning, however, crease patient survival rates. While CNN-based networks
has the potential to revolutionise the area of neurology and are widely utilised for automatic brain tumour segmentation,
enhance patient outcomes with further study and develop- MRI is mostly used for visualising brain tumours. A novel
ment. hybrid weights alignment with a multi-dilated attention
2. LITERATURE SURVEY network (Hybrid-DANet) has been presented to get beyond
the drawbacks of past techniques. To extract high-quality,
Shubhangi Solanki explains that because of the lo-
scale-aware, contextual, and targeted characteristics, this
cation, shape, and size of brain tumours, detection can
network uses a unit of subjects on a traditional encoder-
be difficult. To detect brain tumours and cancers through
decoder function [9]. Performance of the Hybrid-DANet
MRI images, researchers have proposed employing artificial
is comparable to that of cutting-edge techniques, and the
intelligence and statistical image depuration approaches.
authors intend to use transformers in subsequent work to
CNNs are among the machine grabbing techniques that
increase accuracy even more.
have been utilised for categorization and have proven to be
the most accurate. Metrics such as dependability, accuracy, Sohaib Asif explains about establishing an accurate and
and computation time should be taken into account to effective deep researching and transfer lore system for auto-
enhance the system’s performance [6]. By utilising several matically diagnosing brain tumours based on MRI data. To
MRI imaging modalities, such a system can aid in the get steep attributes from the MRI scans, the study used pre-
development of diagnostic tools for a divergence of brain learned exemplars like DenseNet121, Inception ResNetV2,
anarchy like Alzheimer’s infection, Parkinson’s distemper, NasNet Large, and Xception. On the MRI-large dataset, the
dementia, and stroke. For better brain tumour identification suggested CNN model with the Xception and the ADAM
and classification, advanced research involving various deep optimizer has the greatest values for definiteness, reactivity,
learning algorithms, such as deep hybrid learning, may be correctness, particularity, and F1-score [9]. The new method
done in the future. outperformed previous models in terms of performance,

https:// journal.uob.edu.bh
70 Shanmuga Sundari M, et al.: Brain Tumor Prediction using Fine-tuned XceptionNet

highlighting the potential for employing deep learning to Mohammad Ashraf Ottom discusses Znet, an innovative
quickly identify brain tumours from MRI data. To improve method for particularizing brain tumours in 2D MR carbon
the accuracy of the system, future study might make use of copies using deep neural openwork and data reforming tech-
bigger datasets and other deep learning approaches. niques, is presented. The technique uses data amplification,
skip-connection, and encoder-decoder designs to spread the
Gazi Jannatul Ferdous introduces LCDEiT which is closeness of a fewer units of expertly defined tumours
a deep learning method for classifying brain tumours in to many numbers of ersatz cases. The method produced
medical images. The LCDEiT model is well suited for excellent results for the Matthews Correlation Coefficient,
short datasets with linear computing complexity and is F1 score, pixel definiteness, and mean dice similarity coef-
meant to avoid problems associated to inductive bias and ficient. The suggested technique may be used to 3D brain
parameter dependence. The model uses a teacher- student volumes and other disorders [7]. The work does, however,
framework, with a vision transformer with an external atten- draw attention to the drawbacks of utilising pixel accuracy
tion mechanism acting as the student model and a bespoke as an assessment criterion for allowable apportionment in
gated-grouped intricacy neural mesh acting as the teacher the event of class imbalance in MR image segmentation.
model [10]. In two benchmark memorandums, Figshare and Proposed method shows the promising potential of AI op-
BraTS-21, the proposed LCDEiT model achieves good ac- erations in medical imaging by being used as a technology
curacy and F1-score, demonstrating its promise for medical for auto-partition of tumours in MR pictures. The suggested
imaging-based diagnosis when quick computing is essential. architecture for classifying, extracting, parcellating, and
Future studies, according to the authors, should address predicting the existence of and extent of brain tumours using
problems with reduced sample class misclassification rates multichannel 3D MRI aggregate will also be expanded,
and increase the experimental database in order to increase according to the scientists. Additionally, they intend to
the model’s universality. investigate deep learning techniques to produce ground-
truth labelling and realistic high-dimensional, multimodal
Ankit Vidyarthi explains about the multi-class catego- neuroimaging data.
rization of steep-league malignant intellect tumours using
a CAD system is a novel approach discussed in this study Nur Suriza Syazwany explains in order to segment
paper. The method uses the Cumulative Variance Method MRI brain tumours, this study suggests using a multimodal
(CVM), a feature selection algorithm, to select pertinent fusion mesh with a bi-directional feature pyramid network
features from among six domains. The three classifiers (MM-BiFPN). To capture complex interactions between
K-Nearest Neighbour (KNN), multi-class Support Vector modalities, the MM-BiFPN uses individual encoders for
Machine (mSVM), and Neural Network (NN) are then each of the four tones (FLAIR, T1-weighted, T1-c, and
trained and tested using the chosen features to estimate T2-weighted). Multi-modal features and multi-scale features
the accuracy of multi-class classification. With an average are combined in the Bi-FPN layer [11]. The robustness of
accuracy of 95.86% utilising the [11] NN classifier on a the suggested strategy was demonstrated by the model’s
real-world dataset of five kinds of malignant brain tumours, performance evaluation using the MICCAI BraTS2018 and
the suggested method surpassed the existing approaches. MICCAI BraTS2020 data files, which showed comparable
The invention of the CVM algorithm and experiments results even with missing modalities. Future research will
with it utilising a multi-class brain imaging dataset and focus on overcoming the model’s computational and time
benchmark models in the machine learning environment consumption restrictions to create more accurate brain tu-
are important contributions of this work. The work may be mour segmentation models.
expanded in the future with the inclusion of deep learning
models for multi-class brain tumour classification and the Muhammad Rizwan discusses in this research, a Gaus-
addition of more photos to the dataset. sian Convolutional Neural Network (GCNN) technique for
detecting and classifying brain tumours (BTs) into menin-
Ling Tan about the multimodal brain tumour image gioma, glioma, and pituitary, as well as differentiating
segmentation manner purported in this article is depends on between glioma grades [13] (Grade-two, Grade-three, and
the ACU-Net network. The technique uses deep separable Grade-four), is presented. The suggested method uses a
convolutional layers, residual skip connections, and an sixteen-layer GCNN model for output class categorization
active contour model to address issues like hazy tumour and overfitting prevention, coupled with Gaussian image
region boundaries, diseased tissue heterogeneity, and picture filtering, CLF Layer, SFT, FC, and dropout layers. Applying
noise. The suggested method performs superior [12] to data augmentation improves the outcomes. On two datasets,
previous algorithms in Dice, Recall, and Precision metrics the suggested technique achieves high accuracy rates prov-
while segmenting images of brain tumours. Although the ing its suitability for BT multi-class categorization.
ACU-Net model is currently only partially adaptable, future
study will involve extending it to more organs for picture Ahmed S. Musallam explains in this delving, a unique
segmentation studies. For the sake of clinical diagnosis, Deep Convolutional Neural Network (DCNN) architecton-
analysis, and treatment, this study is important. ics is presented for the precise diagnosis of normal brain
pictures, gliomas, meningiomas, and pituitary tumours in

https:// journal.uob.edu.bh
 Int. J. Com. Dig. Sys. 15, No.1, 67-79 (Jan-24) 71

MRI scans. The suggested architecture uses batch normali- Supertwisting. Lyapunov theory is used to assess the perma-
sation for quicker training and simple weight initialization nence of the authority, and MATLAB/Simulink simulations
together with a three-step preprocessing strategy to enhance are run to compare the controllers [17]. The mirroring
image quality [14]. A dataset of 3394 MRI scans used in the consequences show that in details of convergence rate,
study’s experimental results shows high accuracy for normal decreased chattering, and lower medication dosage, the
pictures. The suggested model is a reliable and efficient Supertwisting controller transacts exceptional than the other
computational tool for MRI image-based automated brain controllers. The Supertwisting controller is advised for the
abnormality detection. chemotherapeutic treatment of brain tumours based on the
results of the simulation.
Kuankuan Hao suggests a novel technique for segment-
ing brain tumours using Convolutional Neural Networks Marwa Ismail discusses on identifying effect in glioblas-
(CNNs) termed generalised pooling (GP) with adaptive toma (GBM), a particularly aggressive brain tumour, in line
weights. Particularly in small object tissues like brain with the tumour field effect notion, which contends that
tumours, conventional pooling techniques like maximum cancer has an effect outside of the borders of the visible
pooling and [18] average pooling frequently lose signif- tumour. Using Co-occurrence of Local Anisotropic Gradient
icant features. The GP approach improves segmentation Orientations (COLLAGE), we developed a thorough MRI-
performance by combining maximum pooling with average based legend called r-DepTH [18] that combines measure-
pooling. It does this by computing adaptive weights within ments of tissue deformations in normal brain tissue brought
a amalgamate kernel based on the intake photographs or on by tumour mass effect with morphological features
feature outline. The experiment impacts show that GP is located within tumour boundaries. We evaluated r-DepTH
powerful in segmenting brain tumours, and it may be for survival risk stratification in three different GBM patient
applied as a generic pooling technique for various CNN- groups and obtained encouraging results, outperforming
based tasks. Its usefulness in various applications may be clinical factors and radiomic/deep-learning characteristics
further investigated in future studies. exclusively from tumour boundaries. Future research will
take into account the direction of tissue deformation and
Hasnain Ali Shah explains the EfficientNet-B0 deep confirm the efficacy of r-DepTH in bigger cohorts and
learning model was improved in this study to categorise prospective studies of different solid tumours.
and find brain tumours in MRI data. With a high total
accuracy of 98.87%, the model outperformed other cutting- Saif Ahmad discusses on the use of transfer learning
edge models. Given the time-consuming nature of manual techniques based on deep learning to identify brain tumours
identification, the study emphasised the necessity of auto- in 2D Magnetic Resonance (MR) pictures. The research
mated diagnostic tools for identifying brain tumours in MRI looked at five conventional classifiers and seven pre- learned
scans [15]. Other deep convolutional neural network models exemplary, including VGG-16, VGG-19, ResNet50, Incep-
will be investigated in more detail, as well as expand- tionResNetV2, InceptionV3, Xception, and DenseNet201.
ing the training dataset. Future research will incorporate With the aid of 10-fold cross-nod and a number of fruition
the suggested technique with additional medical imaging indicators, including definiteness, sureness, anamnesis, F1-
modalities. score, Cohen’s kappa, AUC, Jaccard, and relevance, the
study assessed the effectiveness of these models [19]. With
Amran Hossain suggests using the YOLOv3 deep neural an accuracy the top model, VGG-19-SVM, outperformed
fretwork model and an electromagnetic (EM) imaging de- earlier research using machine learning models for brain
vice to find brain tumours. Images are reconstructed using tumour diagnosis. Future study, according to the authors,
a modified delay-multiply-and-sum method after scatter- should evaluate the strategy using several MRI modalities
ing parameters are gathered using a nine-antenna batch and other imaging methods as well as extending it to cate-
configuration with a calico-imitating head hallucination. gorise various tumour kinds. Larger datasets and enhanced
For learning, acceptance and examining, a dataset of 1000 GPU processing may also result in further gains in precision
pictures is prepared, containing 50 specimens with lone and and calculation speed.
dual tumours [16]. The obtained F1 ratings, indicate great
accuracy in terms of detection. The YOLOv3 architecture M. V. S. Ramprasad the BTFSC-Net, a revolutionary
demonstrates its promise for locating and identifying brain medical diagnosis tool that uses artificial intelligence to
tumours utilising a portable EM imaging device by exhibit- categorise brain tumours. The method entails a number
ing high upbringing and acceptance definiteness with low of steps, including preprocessing medical images with a
seasoning and corroboration loss. hybrid probabilistic Wiener filter, combining MRI and CT
carbon copies using a DLCNN with strong edge resolu-
Muhammad Zubair explains that, In order to manage tion stage settings, segmenting the affected region using
the amount of chemotherapy medication given to patients a hybrid fuzzy c-means interspersed k-means heed, and
with brain tumours, this research proposes four distinct extracting hybrid features like coarseness, tint, and sunken-
variable structure controllers, namely Sailing Custom, In- aligned countenance[20]. Finally, a DLPNN is employed
tegral Sailing Mode, Double Integral Sliding Custom, and to differentiate between obliging and pestilential tumours.

https:// journal.uob.edu.bh
72 Shanmuga Sundari M, et al.: Brain Tumor Prediction using Fine-tuned XceptionNet

The BTFSC-Net outperformed existing techniques with also comprises feature extraction based on fineness, earnest-
high segmentation accuracy and classification accuracy. ness, and silhouette, character filtering, and categorizing.
The suggested approach demonstrates real-time application Utilising a variety of performance indicators, the suggested
potential and could be improved further with cutting- edge method achieves a consistent accuracy for pronouncement
optimisation techniques for thorough feature extraction. case history of brain tumours and cataloguing essential facts
of intent stroke [24]. The suggested strategy successfully
C. Yan proposes the use of SEResU-Net, an enhanced divides cases of acute and sub-acute strokes as well as
U-Net model, to automatically segment brain tumours from lofty-calibre and shallow-calibre tumours in brain. The
MRI data. In order to extract additional feature information experimental study shows that the suggested strategy works
and avoid information loss, the model mixes squeeze-and- better than current classification techniques.According to
excitation nexus and deep surplus networks. In order to the article, deep neural networks might be used to detect
address the difficulties of tracery concurrence and figures a variety of brain diseases in the future by gathering input
disproportion, a fusion markdown service is also imple- photos from different databases without the requirement
mented [21]. The model performs better than current pio- for scoop magnification to effectuate soaring results and
neering models in tests on the BraTS2018 and BraTS2019 subside delusion in disease prediction.
compilations. The model is a 2D network, while future
research might take into account utilising a 3D network A. Kujur discusses in this study, the effectiveness of four
architecture to more effectively use the 3D information CNN models—S-CNN, ResNet50, InceptionV3, and Xcep-
present in MRI data and boost segmentation accuracy. tion—was assessed using PCA and two sets of brain MRI
image datasets for intellect tumour and Alzheimer’s disease,
N. Micallef explains about, In order to segregate respectively. Aim was to look into how data complexity
brain tumours, the study offers a new iteration of the affected how well brain MRI predictive models performed.
U-Net++ model. High Dice Coefficient scores were ob- The results demonstrated that the data [25] complexity had
tained using the proposed strategy for the BraTS 2019 a substantial impact on the conduct of the CNN paradigms,
challenge’s Validation Dataset. Numerous modifications to with slighter miscellaneous notes producing greater scores
the U-Net++ model was made, including changes to the than more intense proofs. According to study, more MRI
calamity province, the quantity of kink buildings, and the pictures of various disorders could be used in future tests
intense surveillance technique. Using strategies for history to get more generalised findings.
heightening and post-rarefaction also increased the model’s
accuracy [22]. The study does, however, admit several S. Montaha explains for better odds of survival, brain
drawbacks, including the extensive training period and tumours must be found early. 3D MRI is frequently utilised
individual expenses. The authors offer a number of potential for tumour investigation. This uses three BraTS knowledge
future advancements, such as more pre-processing steps, and information to differentiate intellect tumours into two
ensembling, and specific networks or pathways for various groups using a hybrid model termed TD-CNN-LSTM. The
samples. prototype, which mixes 3D CNN and LSTM, is created
with the best configuration possible and performs well in
K. Venkatachalam explains about a Content-Based Med- studies of layer architecture and hyper-parameter ablation.
ical Image Retrieval (CBMIR) custom can be used to Each MRI sequence is also used to train a 3D CNN model
find images of brain tumours in a sizable MRI image to compare performance [30]. With a test accuracy, the
library. To provide accurate and reliable image retrieval, TD-CNN-LSTM network exceeds the 3D CNN model,
the suggested method makes use of a feature extraction according to the documents. K-fold cross-evidences are
strategy that includes Gabor filtering, the Walsh- Hadamard worn to verify the model’s robustness and consistency of
transform, Fuzzy C-Means clustering, and the Minkowski performance across many training circumstances. In the
hinterland rhythmic [23]. The chain reaction of the tryout future, radiologists may be able to diagnose brain tumours
shows that the proffered brainchild outplays extant methods more accurately thanks to this method of combining a CNN
in details of definiteness and productivity. In terms of model with all MRI sequence analysis.
separating false positive photos and high pixel similarity,
the method has some drawbacks. Future work on these 3. PROPOSED SYSTEM
problems could make use of optimisation algorithms and Detecting brain tumors using XceptionNet is a chal-
semantic-based similarity computation methods. In table lenging problem in medical image analysis. Here’s a brief
explains the literature survey and important future works outline of a proposed system for detecting brain tumors
in the previous research. using XceptionNet:

B. Deepa discusses in this study, a automaton learning A. Data Collection

outcross method is volunteered for categorising disorders The first step is to rendezvous a memorandum of ap-
of the brain, such as tumours and strokes. A hybridised pearances for that we have used Br35H :: Brain Tumor
support cornerstone intended desultory growth classifier is Detection 2020 from Kaggle [31] which consists of binder:
utilized to classify data using the suggested method, which unquestionable and questionable which contains 3000 Brain
MRI Images of size 88MB in which the wrapper amen
https:// journal.uob.edu.bh
 Int. J. Com. Dig. Sys. 15, No.1, 67-79 (Jan-24) 73

TABLE I. FUTURE WORK OF FEW REFERENCE PAPERS

Ref. Citation Future Work

Putting this concept into practise with a variety of deep learning algorithms
1 as deep hybrid learning for detecting and classifying brain tumours will be
additional research for this study.
It is possible to enhance the correctness of brain tumor summary through
MRI by engrossing larger datasets and deep learning approaches such as
5
GAN. These techniques have the potential to enhance the performance of the
diagnosis process.
Enhancing the suggested CNN architecture’s accuracy and effectiveness in
12
MRI image-based brain tumour detection.
Examining the use of extra pre-processing procedures to detect brain tumours
14 that are more accurately and robustly, such as image registration, normalisa-
tion, and segmentation.
Evaluating how the suggested approach performs on datasets with various
18 levels of noise, artefacts, and imaging protocols to gauge how sensitive it is
to changes in picture quality and acquisition conditions.

contains 1500 Brain MRI carbon copy that are carcinoma

and the portfolio questionable contains 1500 Brain MRI
similitude that are non-tumefaction.
B. Data Preprocessing
Before schooling the facsimile, it is essential to pre-
process the experiment to ensure that it is in a commodious
makeup for the Xception net. Step includes resizing the
images to a standard size. Common data augmentation
techniques include resizing.
C. Model Training and Testing
The data has been chinked into a 60% skilled set and
a 40% trying out set for Xception net. During training, Figure 2. Architecture of the Proposed System
the model learns to classify images [32] by adjusting its
weights. Once trained, the model is evaluated on the testing
set to measure its performance. Xception is convolutional semantic criss-cross (CNN)
advanced by François Chollet in 2016. It is a variant of
D. Performance Metrics the Inception architecture known for its ability to capture
The performance metrics used are precision, recall, diverse features from images. With 71 layers, Xception
F1-score and accuracy which measures the percentage of is a deep neural network that has been pretrained on a
correctly classified images. massive statistic of over a million images from the Ima-
geNet database. This training has enabled Xception to learn
E. Prediction rich feature representations for a wide range of objects,
Finally, after cultivation and testing the model, it can including keyboards, mice, pencils, and animals, among
be used to predict and classify the images. To make a others [33]. As a result, the pretrained Xception network is
prediction, the input replica is fed into the experienced capable of accurately classifying images into 1000 object
quintessential, and the output is the predicted label. categories and can be passed down for sundry mainframe
perception millstone.
Since 2014. Additionally, unlike Inception, this does not
start ball rolling any non-straight after the first operation, In Xception, extreme inception architecture takes the
making it different in this aspect as well. principles of Inception to a new level. Unlike Inception,
which uses 1x1 convolutions to reduce input dimensionality
In addition to depthwise separable convolutions, Xcep- and then applies different filters to each of the reduced
tion also employs residual connections, which are skip input spaces, Xception reverses this process. It applies filters
connections that allow gradients to flow more easily during to each of the depth maps first and then compresses the
training and help mitigate the vanishing gradient problem. input space using 1x1 convolutions applied across the depth.
This approach is similar to a depthwise divisible helix, a

https:// journal.uob.edu.bh
74 Shanmuga Sundari M, et al.: Brain Tumor Prediction using Fine-tuned XceptionNet

technique used in neural network design performance of 5. MATHEMATICAL FUNCTIONS

Xception by allowing for deeper networks with improved Mathematical Functions used in the Xception architec-
accuracy. ture and it explains how the inputs are passed to the next
layer.
4. ARCHITECTURE
The Entry Flow, Middle Flow, and Exit Flow are the A. Seperation Convolution
three main sections of the Xception architecture. The separable convolution is a combination of depth
wise sinuosity and point- wise involution. Mathematical
A. Entry Flow
formula for separable convolution is as follows:
The initial feature extraction from the input image is
H,W
done by the Entry Flow. Following two sets of Separable- X
Conv2D layers, it has a folio of curlicue flag and pooling yi , j, K = xi+h−1, j+w−1,m ∗ Wh,w,m,k (1)
mantle. The second set has a window stature of 3x3 and a h,w
parade of 2, whereas the first set uses multiple layers with
a map size of 3x3 and a tramp of 1. The feature maps are In equation 1 x is the proposal tensor, y is the gain
downsampled while the number of channels is increased tensor, w is the burden tensor, H and W are the height and
using these two sets of SeparableConv2D layers. width of the filter, and m is the channel index.
B. Middle Flow B. Batch Normalization
The majority of feature extraction takes place in the Batch normalization is used to normalize the input
Middle Flow. Each of its repeating residual modules has to each layer, so that the distribution of each feature is
three SeparableConv2D layers with the same number of roughly the same. The mathematical formula for batch
channels. It is made up of a succession of these modules. normalization is as follows:
The vanishing gradient issue during training is helped by the x − E[x]
residual connections, which enable gradient propagation. y= √ ∗γ+β (2)
var[x] + ε
C. Exit Flow
The last feature extraction and categorization are done In equation 2 x is the forewarning tensor, E[x] and
by the Exit Flow. A international correctly put together Var[x] are the penurious and variety of the input, epsilon is
layer, a thoroughly coherent ply with 2048 units, and a a small constant, and gamma and beta are enabling vicinity.
ReLU activation function are the layers that make up this
C. ReLU Activation
system. A utterly undivided seam with units equal to the
units of groups in the data and a softmax brisk mission ReLU (Rectified Linear Unit) activation is a non-straight
make up the latter classification flap. exhilarant province that sets all negative values to zero. The
mathematical formula is:
Overall, the Xception architecture incorporates residual
connections to aid in the training of deep neural networks y = max(x, 0) (3)
and performs efficient and precise feature extraction with
depthwise separable convolutions. In equation 3 x is the teaching tensor and y is the crop
tensor.
D. Depthwise Separable Convolution
It is similar to the separable convolution, but the
pointwise convolution is replaced with a depthwise whorl
successive by a pennywise helix. The mathematical formula
for depthwise separable entanglement is as follows:
H,W
X
yi, j,k = xi+h−1, j+w−1 ∗ Wh.m.w ∗ γk βk (4)
h,w

In equation 4 x is the tidings tensor, y is the production

tensor, w is the crammed tensor, H and W are the height
and width of the filter, m is the channel index, and gamma
and beta are learnable parameters.
Figure 3. Architecture of the Xception Net Image Source E. Global Average Pooling
Global average pooling is utilized to diminish the spatial
measurements of the hallmark maps to a single value
https:// journal.uob.edu.bh
 Int. J. Com. Dig. Sys. 15, No.1, 67-79 (Jan-24) 75

per channel. The mathematical formula for global average combining them across channels. This helps to mix the
pooling is as follows: spatial features learned in the depthwise convolution across
H,W different channels, allowing for more complex and abstract
1 X
representations to be learned.
yk = ∗ xi, j,k (5)
H ∗ W i, j
6. XCEPTIONNET ALGORITHM

In equation 5 x is the consultation tensor, y is the turnout Algorithm 1: XCEPTIONNET ALGORITH

tensor, H and W are the height and expanse of the help Result: • Xception model xception base =
tensor, and k is the channel index. keras.applications.Xception
(include top=includet op, weights =
F. Fully Connected Layer weights, input shape =
It is utilized to map the output of the universal standard input shape, pooling = pooling)
uniting to the desired output size. Mathematical formula for initialization;
a fully connected layer is as follows: include top: Dichotomous, whether to comprise the
extensively-tetter layer at the pinnacle of the
y = Wx + b (6) meshwork or not ;
weights: String, pre-training weight to be loaded;
In equation 6 x is the input tensor, W is the tonnage input shape: Tuple of integers, the input shape of the
mold, W is the tonnage mold, b is the unfairness bearing, images in (height, width, channels);
and y is the out-turn tensor. In Xception architecture, these pooling: String, the type of pooling to be applied to
mathematical functions are combined in various ways to the last convolutional layer output classes: Integer,
make a extensive expert system with high accuracy in the number of output classes;
image classification tasks. Intake tensor is slides through a while the model starts do
procession of convolutional layers with group organization Loop: for layer in xception base.layers:
and ReLU activation, followed by a motion of depthwise layer.trainable = False x = xception base.output if
differentiable complication layer with batch formalization keras meets the requirements then
and ReLU activation. Finally, the result of the convolutional x = keras.layers.Flatten()(x);
layers [26] is acknowledged through a overall norm stacking x = keras.layers.Dense(256,
level and a completely inter connected mesh to produce the activation=’relu’)(x) x =
final classification output. keras.layers.Dropout(0.5)(x);
predictions = keras.layers.Dense(classes,
Xception has been pretrained on a large knowledge of activation=’softmax’)(x);
over a million resemblance from the ImageNet database, model =
which enables it to learn rich feature representations for a keras.models.Model(inputs=xception base.input,
wide range of objects. outputs=predictions) ;
return model
Xception has been widely handed-down in analogy else
vision applications, including portrayal classing, target spot- end
ting, image severance, and transfer learning for other visual
recognition tasks. It has achieved futuristic accomplish- The formulas used in algorithms describe the prediction
ment in correctness and computational efficiency in many process using python keras. Above algorithm defines a
benchmark datasets and competitions, making it a popular Keras model named xception model that uses the Xcep-
choice in the research and industry communities for various tion architecture as a base model. The Xception prototype
visual recognition tasks. Depthwise detachable serpentine is is a profound CNN that is disciplined on the ImageNet
composed of two main components: dataset. The xception model function takes several input
parameters including include-top which is a boolean value
G. Depthwise Gyration indicating positively to encompass the utterly-fused surface
In this step, a separate convolutional filter is applied to at the brink of the lattice or not, poundage which is a
each input channel independently. This is typically done string representing the pre-training weights to be loaded,
using small filters, such as 3x3, to capture local spatial input-shape which is a tuple of integers representing the
patterns within each channel. This results in a set of output dimensions of the input image, pooling which is a string
feature maps, one for each input channel. indicating the type of pooling to be applied to the last
convolutional layer output, and classes which is an integer
H. Pointwise Convolution indicating the number of output classes.
After the deepness coil, a 1x1 helix (also known as a
pointwise convolution) is appertaining to the deliverable’s The function starts by instantiating the Xception base
commentary maps. The 1x1 filters are bid to the output model with the given input parameters. It then freezes all
hallmark delineate from the depthwise convolution, linearly the layers of the base model to prevent them from being
https:// journal.uob.edu.bh
76 Shanmuga Sundari M, et al.: Brain Tumor Prediction using Fine-tuned XceptionNet

updated during training.Next, the function adds a custom that the model or system projected as positive is known as
top layer to the model consisting of a flatten layer, a murky precision and the precision of the project is shown below:
layer with 256 units and a ReLU activation concomitant, a TP
eccentric film with a rate of 0.5, and a impenetrable output Precision = (7)
veneer with softmax activation for multi-class stratification T P + FP
[27]. Finally, it fabricate a Keras Model object that takes
the input tensor of the Xception base model and outputs the
predictions of the custom top layers.This algorithm provides
a convenient way to create a transfer learning model for
image classification tasks using the Xception architecture
with a custom top layer. The following algorithm used for
Xceptionnet in our model.
7. EXPERIMENTAL ANALYSIS
We have collected the dataset of images for that we have
used Br35H :: Brain Tumor Detection 2020 from Kaggle
which consists of brochure: referendum division of which
contains 3000 Brain computer assisted tomography images
of size 88MB in which the sub division of certainty contains
1500 images that are cancerous growth and the reverse
binder 1500 Brain MRI Images that are non-malignant
Figure 6. Graph representing the precision of the model
growth. Later we have augumented the images by resizing
its shape.
In machine learning, ”precision” is a metric used to eval-
8. RESULT ANALYSIS uate the performance of a classification model, particularly
In the Fig. 5 given below is the set of images for which in situations where the goal is to minimize false positives.
the model predicted as YES indicating the presence of It is one of the components of the confusion matrix, which
tumor. In the Fig. 6 given below is the set of images for is a table that visualizes the performance of a classification
algorithm.
B. Recall
TP
Recall = (8)
T P + FN

In equation 8, Recall is the segment of true positives

to the total number of instances that were expected to be
Figure 4. Model predicting YES positive and the recall of the project is shown below:

which the model predicted as NO indicating the absence

of tumor. These are the resulted ratios of training and

Figure 5. Model Predicting No

validations with regard to different metrics: TP = True

Positive, FP = False Positive, TN = True Negative, FN =
False Negative
Figure 7. Graph representing the Recall of the model
A. Precision
In equation 7, the ratio of true positives (occurrences that
were accurately recognized) to the total number of instances

https:// journal.uob.edu.bh
 Int. J. Com. Dig. Sys. 15, No.1, 67-79 (Jan-24) 77

C. Accuracy
TP + TN
Accuracy = (9)
T P + T N + FP + FN

In equation 9, The proportion of instances in the dataset

to total right predictions (true positives and true negatives)
is known as accuracy and the recall of the project is shown
below. In machine learning, ”accuracy” is a metric used

Figure 9. Graph representing the F1-Score of the model

Figure 8. Graph representing the accuracy of the model

to evaluate the performance of a classification model. It

measures the ratio of correctly predicted instances to the
total instances in the dataset. Accuracy is one of the most
straightforward and commonly used metrics for classifica-
tion problems.
Figure 9 illustrates the accuracy, demonstrating that the Figure 10. Graph representing the loss of the model
model achieves a 94% success rate. This indicates the
model’s superior performance in disease prediction.
9. CONCLUSION AND FUTURE WORK
D. F1 Score The utilization of XceptionNet as a deep learning
2 ∗ Precision ∗ Recall architecture has exhibited successful results in detecting
F1S core = (10) brain tumors from MRI images. The model achieves 94%
Precision + Recall accuracy indicating its potential as a valuable tool for as-
sisting specialists, doctors in the detection of brain tumors.
In equation 10, F1 score is a single score that balances Additionally to improve the performance of the model, inte-
exactness and recall. It is the consonant mean of precision grating additional data sources like patient medical records
and recall and the F1 score of the project is given below: and imaging modalities can improve accuracy and decrease
E. Model Loss false positives. The interpretability of the model can also
be enhanced by implementing visualization techniques to
L(y true, y pred) = −A ∗ B (11) better understand the highlighted features and brain regions.
Future studies may explore transfer learning methods, which
A = (y true ∗ log(y pred) + (1 − y true)) (12) involve optimizing pre-trained XceptionNet models for spe-
cific brain tumor detection tasks. This approach can lead to
B = log(1 − y pred) (13)
better performance on smaller datasets that are commonly
In equation 11 to 13, where y true is the true label and used in medical imaging applications. Overall, the use of
y pred is the predicted odds of the positive class and XceptionNet in brain tumor discernment has significant
the model loss is shown below Using Xception Net, the future in improving the fidelity and efficacy of brain tumor
model can predict and classify brain MRI scan images pinpointing and treatment.
with accuracy 94%. The above illustrations represent the
progress of model over 13 epochs.
https:// journal.uob.edu.bh
78 Shanmuga Sundari M, et al.: Brain Tumor Prediction using Fine-tuned XceptionNet

R EFERENCES [15] H. A. Shah, F. Saeed, S. Yun, J.-H. Park, A. Paul, and J.-M. Kang,
“A robust approach for brain tumor detection in magnetic resonance
[1] G. Raut, A. Raut, J. Bhagade, J. Bhagade, and S. Gavhane, “Deep
images using finetuned efficientnet,” IEEE Access, vol. 10, pp.
learning approach for brain tumor detection and segmentation,” pp.
65 426–65 438, 2022.
1–5, 2020.
[16] A. Hossain, M. T. Islam, M. S. Islam, M. E. Chowdhury, A. F.
[2] S. Gupta and M. Gupta, “Deep learning for brain tumor segmen-
Almutairi, Q. A. Razouqi, and N. Misran, “A yolov3 deep neural
tation using magnetic resonance images,” in 2021 IEEE conference
network model to detect brain tumor in portable electromagnetic
on computational intelligence in bioinformatics and computational
imaging system,” IEEE Access, vol. 9, pp. 82 647–82 660, 2021.
biology (CIBCB). IEEE, 2021, pp. 1–6.
[17] M. Zubair, I. A. Rana, Y. Islam, and S. A. Khan, “Variable structure
[3] S. Solanki, U. P. Singh, S. S. Chouhan, and S. Jain, “Brain
based control for the chemotherapy of brain tumor,” IEEE Access,
tumor detection and classification using intelligence techniques: An
vol. 9, pp. 107 333–107 346, 2021.
overview,” IEEE Access, 2023.
[18] M. Ismail, P. Prasanna, K. Bera, V. Statsevych, V. Hill, G. Singh,
[4] A. Bs, A. V. Gk, S. Rao, M. Beniwal, and H. J. Pandya, “Electrical
S. Partovi, N. Beig, S. McGarry, P. Laviolette et al., “Radiomic
phenotyping of human brain tissues: An automated system for tumor
deformation and textural heterogeneity (r-depth) descriptor to char-
delineation,” IEEE Access, vol. 10, pp. 17 908–17 919, 2022.
acterize tumor field effect: Application to survival prediction in
glioblastoma,” IEEE transactions on medical imaging, vol. 41, no. 7,
[5] E. Klint, S. Mauritzon, B. Ragnemalm, J. Richter, and K. Wrdell, pp. 1764–1777, 2022.
“Fluora-a system for combined fluorescence and microcirculation
measurements in brain tumor surgery,” in 2021 43rd Annual Inter-
[19] S. Ahmad and P. K. Choudhury, “On the performance of deep
national Conference of the IEEE Engineering in Medicine & Biology
transfer learning networks for brain tumor detection using mr
Society (EMBC). IEEE, 2021, pp. 1512–1515.
images,” IEEE Access, vol. 10, pp. 59 099–59 114, 2022.
[6] N. Ilyas, Y. Song, A. Raja, and B. Lee, “Hybrid-danet: an encoder-
[20] M. Ramprasad, M. Z. U. Rahman, and M. D. Bayleyegn, “A
decoder based hybrid weights alignment with multi-dilated attention
deep probabilistic sensing and learning model for brain tumor
network for automatic brain tumor segmentation,” IEEE Access,
classification with fusion-net and hfcmik segmentation,” IEEE Open
vol. 10, pp. 122 658–122 669, 2022.
Journal of Engineering in Medicine and Biology, vol. 3, pp. 178–
188, 2022.
[7] S. Asif, W. Yi, Q. U. Ain, J. Hou, T. Yi, and J. Si, “Improving
effectiveness of different deep transfer learning-based models for
[21] C. Yan, J. Ding, H. Zhang, K. Tong, B. Hua, and S. Shi, “Seresu-net
detecting brain tumors from mr images,” IEEE Access, vol. 10, pp.
for multimodal brain tumor segmentation,” IEEE Access, vol. 10, pp.
34 716–34 730, 2022.
117 033–117 044, 2022.
[8] G. J. Ferdous, K. A. Sathi, M. A. Hossain, M. M. Hoque, and
[22] N. Micallef, D. Seychell, and C. J. Bajada, “Exploring the u-net++
M. A. A. Dewan, “Lcdeit: A linear complexity data-efficient image
model for automatic brain tumor segmentation,” IEEE Access, vol. 9,
transformer for mri brain tumor classification,” IEEE Access, vol. 11,
pp. 125 523–125 539, 2021.
pp. 20 337–20 350, 2023.
[23] K. Venkatachalam, S. Siuly, N. Bacanin, S. Hubálovskỳ, and
[9] A. Vidyarthi, R. Agarwal, D. Gupta, R. Sharma, D. Draheim, and
P. Trojovskỳ, “An efficient gabor walsh-hadamard transform based
P. Tiwari, “Machine learning assisted methodology for multiclass
approach for retrieving brain tumor images from mri,” IEEE Access,
classification of malignant brain tumors,” IEEE Access, vol. 10, pp.
vol. 9, pp. 119 078–119 089, 2021.
50 624–50 640, 2022.
[24] B. Deepa, M. Murugappan, M. Sumithra, M. Mahmud, and M. S.
[10] L. Tan, W. Ma, J. Xia, and S. Sarker, “Multimodal magnetic reso-
Al-Rakhami, “Pattern descriptors orientation and map firefly algo-
nance image brain tumor segmentation based on acu-net network,”
rithm based brain pathology classification using hybridized machine
IEEE Access, vol. 9, pp. 14 608–14 618, 2021.
learning algorithm,” IEEE Access, vol. 10, pp. 3848–3863, 2021.
[11] N. S. Syazwany, J.-H. Nam, and S.-C. Lee, “Mm-bifpn: multi-
[25] A. Kujur, Z. Raza, A. A. Khan, and C. Wechtaisong, “Data
modality fusion network with bi-fpn for mri brain tumor segmenta-
complexity based evaluation of the model dependence of brain mri
tion,” IEEE Access, vol. 9, pp. 160 708–160 720, 2021.
images for classification of brain tumor and alzheimer’s disease,”
IEEE Access, vol. 10, pp. 112 117–112 133, 2022.
[12] M. A. Ottom, H. A. Rahman, and I. D. Dinov, “Znet: deep learning
approach for 2d mri brain tumor segmentation,” IEEE Journal of
[26] M. Shanmuga Sundari, M. Sudha Rani, and K. B. Ram, “Acute
Translational Engineering in Health and Medicine, vol. 10, pp. 1–
leukemia classification and prediction in blood cells using convo-
8, 2022.
lution neural network,” in International Conference on Innovative
Computing and Communications: Proceedings of ICICC 2022,
[13] M. Rizwan, A. Shabbir, A. R. Javed, M. Shabbir, T. Baker, and Volume 1. Springer, 2022, pp. 129–137.
D. A.-J. Obe, “Brain tumor and glioma grade classification using
gaussian convolutional neural network,” IEEE Access, vol. 10, pp.
[27] M. Shanmuga Sundari and V. C. Jadala, “Neurological disease
29 731–29 740, 2022.
prediction using impaired gait analysis for foot position in cerebellar
ataxia by ensemble approach,” Automatika, vol. 64, no. 3, pp. 541–
[14] A. S. Musallam, A. S. Sherif, and M. K. Hussein, “A new con- 550, 2023.
volutional neural network architecture for automatic detection of
brain tumors in magnetic resonance imaging images,” IEEE access,
vol. 10, pp. 2775–2782, 2022.

https:// journal.uob.edu.bh
 Int. J. Com. Dig. Sys. 15, No.1, 67-79 (Jan-24) 79

Shanmuga Sundari.M Shanmuga Sundari Vidyullatha Sukhavasi Vidyullatha

M is a research scholar at K LUniversity; Sukhavasi is a research scholar atVignan’s
Interested Research areas are Machine learn- Foundation for Science,Technology and
ing, Artificial Intelligence, Software Engi- Research. Interested in Deep learning.
neering. Published 25 Publications. Published 7 Publications.

M.Dyva Sugnana Rao M.Dyva Sugnana

Yeluri Divya Ms. Y Divya is work- Rao is a research scholar at JNTUH; Inter-
ing as an Assistant Professor and have 4 ested Research areas are Machine learning,
years of teaching experience. Interested in Data structure. Published 4 Publications.
DBMS,COA,AI and COOS. My research
area of interest is Machine Learning and
Deep Learning.

M. Sudha Rani M.Sudha Rani is a research

KBKS Durga KBKS Durga is work- scholar at JNTUH; Interested Research areas
ing as an Assistant Professor and have 4 are Machine learning. Published 3 Publica-
years of teaching experience. Interested in tions.
DBMS,COA and networks. My research
area of interest is Machine Learning.

https:// journal.uob.edu.bh