Deep Learning based Brain Stroke Detection using Improved VGGNet
SRISABARIMANI K. ARTHI R
Electronics and Communication Engineering,
SRM Institute of Science and Technology,
Ramapuram Campus Chennai Tamilnadu,
INDIA
Abstract: - Brain stroke is one of the critical health issues as the after effects provides physical inability and
sometimes death. The inability of focus in the brain due to bleeding or clogged blood vessels leads to stroke.
Early treatment and diagnosis are crucial in and following manual procedures takes more time which further
increases the criticalness. Artificial intelligence and machine learning techniques hands together in medical
domain and numerous applications are developed to reduce the diagnosis time and to improve the accuracy.
Incorporating machine learning techniques in brain stroke detection is a familiar research arena and numerous
research works are evolved with better solutions. However, the drive towards developing better system for
brain stroke detection is still in progress. Thus, in this research work, deep learning-based brain stroke detection
system is presented using improved VGGNet. Simulation analysis using a set of brain stroke data and the
performance of learning algorithms are measured in terms of accuracy, sensitivity, specificity, precision, f-
measure, and Jaccard index. The better performance of proposed model is comparatively analyzed with
traditional machine learning algorithms like support vector machine, Naïve Bayes, Decision tree, K-Nearest
neighbor, and recent deep learning models like ResNet, Squeeze Net, Alex Net, and Google Net algorithms.
Experimental results validates that the Improved VGG model attained better performance for all the
parameters. Specifically with 96.86% of detection accuracy improved VGG model detects the brain strokes
effectively compared to other learning algorithms.
Key-Words: - Brain stroke detection, Machine Learning, Deep Learning, Convolution Neural Network,
Detection Accuracy
Received: June 27, 2022. Revised: September 19, 2023. Accepted: October 5, 2023. Published: October 12, 2023.
1 Introduction
Stroke is one of the serious medical issues which
require immediate medical attention to avoid further
complications. The major reason for stroke is
improper life style and its impacts in body mass
index, reduced function of heart and kidney,
inappropriate level of glucose, drinking, smoking,
etc., When the blood flow to brain is reduced or
disrupted the cells in the brain reduces its functions
which leads to stroke. Every year fifteen million
people suffer due to stroke as per the report from
world health organization (WHO). Center for
disease control and prevention reports that stroke is
the sixth leading cause of mortality specifically in
India stroke occupies fourth position. Stroke is a
noncommunicable disease and it is mainly classified
into hemorrhagic and ischemic. Hemorrhagic stroke
occurs due to bleeds in blood vessels. In critical
cases, burst blood vessel leads to hemorrhagic
stroke. Ischemic stroke occurs due to blocks in
blood vessels which go to brain. Figure 1 depicts an
illustration for different types of strokes. The
ischemic stroke is further classified into thrombotic
and embolic. Thrombotic occurs when there is block
or clot in artery which provides blood supply to
brain whereas embolic occurs when blocks in any
part of the body and move towards the brain. As of
now there is no medicine or treatment available to
cure the stroke completely however medications are
available to extend the stroke patients lifespan.
Thus, it is essential to predict or detect the
symptoms of strokes from health records.
Fig. 1: Sample stroke MR Images
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2023.20.21
Srisabarimani K. Arthi R
E-ISSN: 2224-2902
204
Volume 20, 2023
Currently magnetic resonance imaging (MRI)
and Computed tomography (CT) scans are widely
used to identify brain diseases. Through MRI the
brain structure and functions can be diagnosed and
the anatomical structure of brain can be observed.
Anomalies in cranial cavity, vertebral column and
inspection of cranial nerves are possible in MRI.
Compared to CT scans the results of MRI are better
and less prone to noise factors. Lesion identification
and segmentation, brain parcellation and tissue
segmentation can be performed using magnetic
resonance images. In numerous medical
applications, artificial intelligence algorithm is used
to classify and detect abnormalities. Specifically,
segmentation and classification of MR images using
machine learning algorithms are increased to detect
health issues.
Machine learning algorithms are like support
vector machine, naïve bayes, decision tree, random
forest algorithms are used to detect stroke type from
brain images. The traditional machine learning
based stroke detection models classifies the
abnormalities based on the features. A separate
feature detection and extraction models are required
for the machine learning algorithms then only the
features can be classified to detect the abnormalities.
However deep learning models comes with
automatic feature extraction and it does not require
any separate model for initial feature selection and
extraction. Various deep learning models are used in
numerous image processing applications. Thus, in
this research work the performance of deep learning
models is analyzed in addition to the proposed
model. The major objective of this research work is
to obtain maximum detection accuracy in brain
stroke detection with minimum false positive rates.
To attain this, an improved VGGNet model is
proposed which effectively detects different types of
strokes. The major contributions of this research
work are presented as follows.
1. An improved deep learning model I-VGGNet is
presented for classifying the brain images to
detect brain stroke types.
2. An intense experimental analysis is presented to
demonstrate the proposed model performance
using different type of stroke and normal brain
images.
3. A detailed comparative analysis is presented
with traditional machine learning models and
recent deep learning models to validate the
proposed model better performance in terms of
accuracy, specificity, sensitivity, precision, f-
measure and Jaccard index.
The remaining discussion in the article is
presented as follows. a brief literature review is
presented in section 2. Proposed improved VGGNet
model is presented in section 3. Detailed
experimental analysis and results are presented in
section 4 and the summary is presented in the last
section.
2 Problem Formulation
2.1 Related Works
The magnetic induction tomography (MIT) brain
pictures are analyzed by the hemorrhagic stroke
detection model that is provided in. Images
generated via induction tomography are analyzed
using the weighted frequency difference adaptive
thresholding split Bregman method, which detects
stroke with the least amount of reconstruction. The
given approach, according to experimental findings,
has a lower reconstruction error than more
traditional multifrequency-based induction
tomography image processing techniques [1]. In all
2D and 3D situations, the proposed fully automatic
MI-UNet stroke lesion segmentation outperformed
UNet. Results generated by 3D MI-UNet have
superior segmentation performance as assessed by
the Dice score, Hausdorff distance, average
symmetric surface distance, and precision [2]. With
the best location accuracy of 0.9859 for detection,
0.8033 Dice score, and 0.6919 IoU for
segmentation, the suggested model achieves
competitive performance with human experts on
two separate datasets. The findings show how the
suggested deep learning model is efficient, reliable,
and advantageous in automatically diagnosing
haemorrhage lesions, making it possible to use it as
a clinical decision support tool for stroke diagnosis
[3]. The bio-signals were collected at a sampling
rate of 1,000Hz per second from the three electrodes
of the ECG and the index finger for PPG while
walking in the suggested system, which takes into
account the convenience of wearing the bio-signal
sensors for the elderly [4]. The goal of this research
is to use the convolutional neural network (CNN)
model to categorize brain CT pictures as normal,
surviving ischemia, or cerebral hemorrhage. We
suggest employing computed tomography images to
classify cerebral strokes using a computer-aided
diagnostic system (CAD). Techniques for horizontal
flip data magnification were applied to gain more
precise categorization. The accuracy and recall of
numerous estimation parameters were both
enhanced by the suggested strategy [5]. This
research employs a deep generative adversarial
network-based method for detecting brain lesions
from brain scans. The proposed method first creates
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2023.20.21
Srisabarimani K. Arthi R
E-ISSN: 2224-2902
205
Volume 20, 2023
fake magnetic resonance images using computed
tomography data. Using the proposed adversarial
network, the lesions are searched from the MR
images and validated against conventional methods
[6]. The brain computer interface model used
electroencephalography and augmented reality to
detect brain strokes. The suggested technique firstly
identifies spatial and spectral features from EEG
signals. The identified features are extracted and
categorized using regularized discriminant analysis.
A kernel density estimation function is also included
to enhance classification performance, providing
more reliable detection results than traditional
approaches [7]. This study proposes novel
automated classification and segmentation
algorithms to simultaneously identify hemorrhage
and ischemic lesions (infarcts) from noncontrast
brain CT images for the treatment of patients with
brain strokes. The goal of the U-Net segmentation
model is to accurately and automatically detect
stroke lesions [8]. The author has suggested a
mechanism for detecting brain tumors using MR
images that is based on fusion. Making use of
machine learning algorithms [9]. To identify the
ischemic stroke lesions, a brand-new, optimized
fuzzy level segmentation approach is put forth in
this study. The multi-textural features are extracted
following segmentation to create a feature set. The
proposed weighted Gaussian Naive Bayes classifier
uses these features as input to distinguish between
classes of normal and pathological stroke lesions
[10]. This study uses Support Vector Machine
(SVM) classifier algorithms and explicit highlight
extraction techniques to identify benign and
malignant brain tumors [11]. By blocking and
adaptively sequencing the convolution layers,
optimizing the number of activation functions and
hyperparameters, and dimensional U-Net (D-UNet)
optimization, a convolutional deep network
architecture is suggested. In order to ascertain
whether there has been a brain stroke, the suggested
method looks at the computed tomography (CT)
pictures from the dataset. Once a stroke has
happened, it is possible to establish whether it was
caused by ischemia or hemorrhage [12]. The author
briefly discussed about five major phenotypes of
stroke occurs via thrombogenic paths [13]. This
article discussed about various segmentation and
classification of tumour images [14].
2.1 Methodology
The proposed deep learning based brain stroke
detection model is presented in this section. The
presented approach incorporates an improved
version of VGGNet to obtain better detection
accuracy. The traditional VGGNet has more layers
and the time required to train the network is high.
Moreover, traditional network performs more
calculations thus it reduces the convergence speed.
Additionally due to large parameters requirement
the memory requirement of traditional network is
high. Thus, an improved VGGNet is presented in
this research work for brain stroke detection. A
complete overview of proposed model is presented
in Figure 2. The input image is initially
preprocessed to remove the noise artifacts using
gaussian filtering. Then the data is fed into network
model to extract the features and classified in the
last layer. The convolution layers in the network
extracts the features and max pooling reduces the
feature dimensions. The final SoftMax function in
the classifier layer classifies the features as normal,
hemorrhage, ischemic strokes.
Fig. 2: Proposed stroke detection model overview
VGGNet is a type of convolutional neural
network which is developed by visual geometry
group of oxford university. Compared to other CNN
models, the feature extraction capabilities of
VGGNet are high and it can able to perform
multiple tasks. compared to Google Net, the
performance of VGGNet will be better in image
processing applications. The architecture of
traditional VGGNet is depicted in Figure 3. The
architecture included series of convolution layers
and max pooling layers. Total five blocks of layers
are present in the VGGNet in which the first two
blocks has two convolution layers and one max
pooling layer. The next three blocks have three
convolution layers in addition to a max pooling
layer. Total 13 convolution layers and 5 pooling
layers are present in the architecture. A flatten layer
is used after last pooling layer to convert the two-
dimensional features into one-dimensional features.
Followed by three fully connected network is used
and finally classification layer classifies the features
using SoftMax function. The kernel size of the
convolution layers is increased for each block.
ReLU activation function is used after convolution
layers and fully connected layers and to avoid data
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2023.20.21
Srisabarimani K. Arthi R
E-ISSN: 2224-2902
206
Volume 20, 2023
overfitting a dropout is used after fully connected
layers.
The convolution layers in the VGGNet extracts
the features from the input. The convolution kernel
in this layer is obtained by convoluting feature map
with series of filters. Mathematically the
convolution process is formulated as
where, k indicates the convolution kernel,
indicates the feature map, represents the input
image and represents the bias term. A non linear
function is generally introduced after convolution
process which introduces non-linearity in the output
feature map and it is mathematically expressed as
where, indicates the output feature map, the
location of feature is indicated as . The term
indicates the input pixel value, weight
and convolution kernel. Followed by convolution
layers, pooling layer is employed for each block
which reduces the feature map dimensions without
removing the essential information. The pooling
function is mathematically expressed as
where indicates the feature map after
pooling operation, indicates the pooling
region. In the proposed VGGNet model, max
pooling is employed so that largest value in the
convoluted feature is selected to obtain new
feature output.
Fig. 3: Architecture of VGGNet and Improved
VGGNet
After final pooling operation, the features are
flattened which provides a one-dimensional feature
vector and it’s processed through three fully
connected layers. after fully connected layers
SoftMax function in the final classifier layer
classifies the features and provides the class
probabilities. Mathematically the SoftMax function
is expressed as
where indicates the detected class, number of
hidden neurons in fully connected layer is indicates
as and weight matrix is indicated as.
indicates the bias term.
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2023.20.21
Srisabarimani K. Arthi R
E-ISSN: 2224-2902
207
Volume 20, 2023
Figure 3 depicts architectures of traditional
VGGNet and improved VGGNet with essential
layer details. It can be seen from the improved
VGGNet layer details the number of parameters is
reduced by reducing the layer depths. This reduces
the possibilities of data overfitting or underfitting
problem in the training process. The first two blocks
in the improved VGGNet and traditional VGGNet
are same and it is used to extract the features. The
consecutive convolutional kernels extract the
features however the input size is changed in the
improved VGGNet. The input size used in the
Improved VGGNet is 128×128 whereas in
traditional model, the input size is 224×224. The
next three blocks in the traditional VGGNet have
four layers in which three are convolution layers and
1 pooling layer. However in the improved VGGNet
it is reduced and the kernel size is changed for each
block as 64, 128 and 64 respectively. This reduced
kernel reduces the parameter requirements, finally
the pooling layer reduces the feature dimensions and
the feature maps are converted into one-dimensional
vectors in flatten layer. Three fully connected layers
in the traditional VGGNet is reduced into two in the
Improved VGGNet model to reduce the parameters
further and finally classified to detect the stroke
type.
3 Results and Discussion
The proposed brain stroke detection model using
learning algorithms is verified through simulation
analysis performed in MATLAB. The dataset used
for the experimentation is prepared manually by
collecting sample images. Total 606 images are
collected in which 186 images are Hemorrhagic, 33
samples belong to Ischemic and the remaining 387
samples are normal samples. The dataset is divided
in the ratio of 80:20 for training and testing purpose
and the details of samples are presented in Table. 1.
Table 1. Data samples used for training and testing
S.No.
Type
Training
Images
Testing
Images
Total
1
Haemorrhagic
149
37
186
2
Ischemic
26
7
33
3
Normal
310
77
387
Total
606
The performance of proposed models evaluated
using the performance metrics like accuracy,
sensitivity, specificity, precision, F-measure and
Jaccard index. Based on the results of confusion
matrix elements like true positive values, false
positive values, true negative values and false
negative values the metrics are calculated.
mathematical expression to calculate the
performance metrics are presented as follows.
Table 2. Proposed model performance metrics
S.No
Performance metrics
Range (%)
1
Accuracy
96.86
2
Sensitivity
96.71
3
Specificity
97.02
4
Precision
97.03
5
F1 Score
96.87
6
Jaccard Index
93.93
Table 2 depicts the performance metrics of
proposed model. The results obtained by the
proposed model are presented for all the metrics. To
validate that the proposed model is better, a detailed
comparative analysis with traditional machine
learning algorithms like support vector machine,
Naïve Bayes, Decision tree, K-Nearest neighbor and
recent deep learning models like ResNet, Squeeze
Net, Alex Net, and Google Net algorithms are
presented. The performance metrics of all the
models are obtained through individual
experimentation and the results are summarized to
present the comparative performance analysis. since
the analysis included 8 models thus instead of
executing all the models combined in simulation
platform those models are trained and tested
individually. For each model the essential parameter
settings are made to obtain the results and
summarized as graphs in Microsoft excel. Figure 4
depicts the comparative analysis of all the models
for sensitivity and specificity parameter.
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2023.20.21
Srisabarimani K. Arthi R
E-ISSN: 2224-2902
208
Volume 20, 2023
Fig. 4: Sensitivity and Specificity analysis
From the results given for sensitivity and
specificity for all the models, it can be observed that
the performance of proposed VGGNet is better than
other models. The sensitivity obtained by the
proposed model is 96.71% which is 21% greater
than the Naïve Bayes algorithm, 19% greater than
the kNN algorithm, 15% greater than the decision
tree algorithm, 13% greater than the SVM
algorithm. When the performance is compared to
deep learning models, the sensitivity obtained by the
proposed model is 7% greater than the Squeeze Net
model, 6% greater than the Alex Net model, 5 %
greater than the Google Net model and 3% greater
than the ResNet model. Similarly for specificity the
proposed model obtained maximum specificity as
97.02% which is 24% greater than the Naïve Bayes,
23% greater than the kNN, 18% greater than the
decision tree, 16% greater than the SVM model.
When specificity is compared with deep learning
algorithms, the proposed model performance is 9%
greater than the Squeeze Net model, 7% greater than
the Alex Net model, 5 % greater than the Google
Net model and 2% greater than the ResNet model.
Figure 5 depicts the comparative analysis of all
the models for precision and F-measure metrics. The
performance of proposed model is maximum for
precision and f-measure and it is clearly visible in
Figure 2 The precision obtained by the proposed
model is 97.03% which is 23% greater than the
Naïve Bayes algorithm, 22% greater than the kNN
algorithm, 18% greater than the decision tree
algorithm, 15% greater than the SVM algorithm.
When the performance is compared to deep learning
models, the precision obtained by the proposed
model is 9% greater than the Squeeze Net model,
7% greater than the Alex Net model, 5% greater
than the Google Net model and 2% greater than the
ResNet model. Similarly for F-measure the
proposed model obtained maximum F-measure as
96.87% which is 22% greater than the Naïve Bayes,
21% greater than the kNN, 16% greater than the
decision tree, 14% greater than the SVM model.
When f-measure is compared with deep learning
algorithms, the proposed model performance is 8%
greater than the Squeeze Net model, 6 % greater
than the Alex Net model, 5% greater than the
Google Net model and 2% greater than the ResNet
model.
Fig. 5: Precision and F1 Score (F-measure) analysis
Fig. 6: Analysis of Jaccard Index
Further the Jaccard index is calculated based on
the f-measure and compared with all the existing
methods. Figure 6 depicts the comparative analysis
of Jaccard index values obtained by the proposed
model and other learning models. The maximum
index value is attained by the proposed model which
is 93.93%. when compared to other models the
index obtained by the proposed model is 34%
greater than the Naïve Bayes algorithm, 32% greater
than the kNN algorithm, 27% greater than the
decision tree algorithm, 23% greater than the SVM
algorithm. When the performance is compared to
deep learning models, the index value obtained by
the proposed model is 14% greater than the Squeeze
Net model, 11% greater than the Alex Net model,
9% greater than the Google Net model and 5%
greater than the ResNet model.
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2023.20.21
Srisabarimani K. Arthi R
E-ISSN: 2224-2902
209
Volume 20, 2023
Fig. 7: Accuracy comparative analysis
Figure 7 depicts the comparative analysis of
detection accuracy obtained by the proposed model
and other learning models. The maximum accuracy
attained by the proposed model is 96.86%. When
compared to other models the accuracy obtained by
the proposed model is 23% greater than the Naïve
Bayes algorithm, 21% greater than the KNN
algorithm, 17% greater than the decision tree
algorithm, 14% greater than the SVM algorithm.
When the performance is compared to deep learning
models, the accuracy obtained by the proposed
model is 8% greater than the Squeeze Net model,
6% greater than the Alex Net model, 5% greater
than the Google Net model and 2% greater than the
ResNet model.
For better understanding details of comparative
analysis are presented as numerical data in Table 3.
From the results given in Table. 2, it can be
observed that the performance of improved VGG
model is better than other models for all the metrics.
Thus, brain stroke can be effectively detected
through the presented improved VGG model.
4 Conclusion
A deep learning-based brain stroke detection model
is presented in this research work using improved
VGGNet. The proposed work extracts the essential
features and detects the type of stroke as
hemorrhagic and ischemic. The detection model
performance is verified through experimentation
using brain stroke dataset and the performances are
compared with machine learning and deep learning
models. The performance analysis considered
traditional machine learning models and recent deep
learning methods for evaluation through the
performance metrics like precision, specificity,
sensitivity, f-measure, precision and Jaccard index.
From the comparative analysis the better
performance of proposed improved VGGNet model
is observed with 96.86% detection accuracy which
is much better than the other deep learning and
machine learning models. Further, this research
work can be extended to attain better detection
performances by incorporating optimization
algorithms with deep learning models to finetune
the parameters of network model.
References:
[1] Tao Zhang, Weirui Zhang, Xuechao Liu,
Meng Dai, Qibo Xuan, Xiuzhen Dong,
Ruigang Liu, Canhua Xu “Multifrequency
Magnetic Induction Tomography for
Hemorrhagic Stroke Detection Using an
Adaptive Threshold Split Bregman
Algorithm”, IEEE Transactions on
Instrumentation and Measurement, Vol. 71,
2022, pp. 1-13.
[2] Yue Zhang, Jiong Wu, Yilong Liu, Yifan
Chen, Ed X. Wu, Xiaoying Tang “MI-UNet:
Multi-Inputs UNet Incorporating Brain
Parcellation for Stroke Lesion Segmentation
from T1-Weighted Magnetic Resonance
Images”, IEEE Journal of Biomedical and
Health Informatics, Vol. 25, No. 2, 2021, pp.
526-535.
[3] Lu Li, Meng Wei, Bo Liu, Kunakorn
Atchaneeyasakul, Fugen Zhou, Zehao Pan,
Shimran A. Kumar, Jason Y. Zhang, Yuehua
Pu, David S. Liebeskind, Fabien Scalzo,
“Deep Learning for Hemorrhagic Lesion
Detection and Segmentation on Brain CT
Images”, IEEE Journal of Biomedical and
Health Informatics, Vol. 25, No. 5, 2021, pp.
1646-1659.
[4] Jaehak Yu, Sejin Park, Soon-Hyun Kwon,
Kang-Hee Cho, Hansung Lee, “AI-Based
Stroke Disease Prediction System Using ECG
and PPG Bio-Signals”, IEEE Access, Vol. 10,
2022, pp. 43623-43638.
[5] Azhar Tursynova, Batyrkhan Omarov,
Natalya Tukenova, Indira Salgozha, Onergul
Khaaval, Rinat Ramazanov, Bagdat Ospanov
“Deep Learning-Enabled Brain Stroke
Classification on Computed Tomography
Images”, Computers, Materials & Continua,
Vol.75, No.1, 2023, pp. 1431-1446.
[6] Na Hu1,Tianwei Zhang, Yifan Wu, Biqiu
Tang, Minlong Li, Bin Song, Qiyong Gong,
Min Wu, Shi Gu, Su Lui, “Detecting brain
lesions in suspected acute ischemic stroke
with CT-based synthetic MRI using
generative adversarial networks”, Annals of
translational medicine, Vol. 10, No.2, 2022,
pp. 1-16.
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2023.20.21
Srisabarimani K. Arthi R
E-ISSN: 2224-2902
210
Volume 20, 2023
[7] Mak J, Kocanaogullari D, Huang X, Kersey J,
Shih M, Grattan ES, Skidmore ER,
Wittenberg GF, Ostadabbas S, Akcakaya M.
“Detection of Stroke-Induced Visual Neglect
and Target Response Prediction Using
Augmented Reality and
Electroencephalography”. IEEE Transactions
on Neural Systems and Rehabilitation
Engineering. Vol, 30, 2022, pp. 1840-50.
[8] Kaya B, Önal M. “A CNN transfer
learning‐based approach for segmentation
and classification of brain stroke from
noncontrast CT images”. International
Journal of Imaging Systems and Technology.
Vol. 33, No. 4, 2023, pp. 1335-1352.
[9] Kaliannan, S., Rengaraj, A. and Daniel, A.P.,
“A Novel Fusion Approach to Detect Brain
Tumor Using Machine Learning for MRI
Images”. Traitement du Signal, Vol. 39, No.
4, 2022, pp. 1363.
[10] Jayachitra, S., and A. Prasanth. "Multi-feature
analysis for automated brain stroke
classification using weighted Gaussian naïve
Bayes classifier." Journal of circuits, systems
and computers Vol. 30, No. 10, 2021, pp.
1793-6454.
[11] Srisabarimani, K., R. Arthi, Nikita Parihar,
Kannu Priya, and Saundrya Nair. "Detection
and classification of different stages of benign
and malignant tumor of MRI brain using
machine learning." In Advances in Power
Systems and Energy Management: Select
Proceedings of ETAEERE, 2021, pp. 319-330.
Springer Singapore.
[12] Sercan Yalçın, Hüseyin Vural “Brain stroke
classification and segmentation using
encoder-decoder based deep convolutional
neural networks” Computers in Biology and
Medicine, Vol 149, 2022, pp.105941
[13] Jae C. Chang, Stroke Classification: Critical
Role of Unusually Large von Willebrand
Factor Multimers and Tissue Factor on
Clinical Phenotypes Based on Novel “Two-
Path Unifying Theory” of Hemostasis,
Clinical and Applied Thrombosis/Hemostasis,
Vol. 26, 2020.
[14] Sri Sabarimani, K., Arthi, R. “A Brief
Review on Brain Tumour Detection and
Classifications”. Bio-inspired
Neurocomputing. Studies in Computational
Intelligence, Vol. 903. 2021, pp. 61-72.
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2023.20.21
Srisabarimani K. Arthi R
E-ISSN: 2224-2902
211
Volume 20, 2023
APPENDIX
Table 3. Overall performance analysis
Method
Accuracy
Sensitivity
Specificity
Precision
F-Measure
Jaccard Index
Decision Tree
80.20
81.46
78.95
79.35
80.39
67.21
Naïve Base
74.26
75.48
72.97
74.52
75.00
60.00
KNN
75.66
77.35
73.91
75.39
76.36
61.76
SVM
82.51
83.61
81.40
81.99
82.79
70.64
ResNet
94.39
94.10
94.68
94.72
94.41
89.41
Squeeze Net
88.94
89.40
88.49
88.52
88.96
80.12
Alex Net
90.59
91.09
90.10
90.20
90.64
82.88
Google Net
91.91
92.11
91.72
91.80
91.95
85.11
Proposed
(I-VGGNet)
96.86
96.71
97.02
97.03
96.87
93.93
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The authors equally contributed in the present
research, at all stages from the formulation of the
problem to the final findings and solution.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
No funding was received for conducting this
research work.
Conflict of Interest
The authors have no conflicts of interest to declare
that are relevant to the content of this article.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2023.20.21
Srisabarimani K. Arthi R
E-ISSN: 2224-2902
212
Volume 20, 2023
