Brain Stroke Prediction Using Visual Geometry Group Model

V. V. L. Narayanan

, A. Reddy

, V. Venkatesh

, S. Tutun

, P. Norouzzadeh

, E. Snir

, S. Mahmoud

and B. Rahmani

1,*

Saint Louis University, Computer Science Department, St. Louis, MO, U.S.A.

Washington University in Saint Louis, Olin Business School, St. Louis, MO, U.S.A.

Saint Louis University, Medical School, St. Louis, MO, U.S.A.

Abstract: Stroke has become the leading cause of high mortality and disability rates in the modern era. Early detection

and prediction of stroke can significantly improve patient outcomes. In this study, we propose a deep learning

approach using the Visual Geometry Group (VGG-16) model. VGG-16 is a type of Convolutional Neural

Network (CNN) which is one of the best computer vision models to date to predict the occurrence of a stroke

in the brain. VGG-16 is a type of CNN that is one of the best computer vision models to date. We used a

dataset consisting of Magnetic resonance imaging (MRI) images of patients with and without stroke. The

VGG-16 model was pre-trained on the ImageNet dataset and fine-tuned on our dataset to predict the

occurrence of a stroke. Our experimental results demonstrated that the proposed approach achieves high

accuracy and can effectively predict stroke occurrence. We have also conducted an extensive analysis of the

model's performance and provided insights into important features used by the model to predict stroke

occurrence. The proposed approach has the potential to be used in clinical settings to aid in the early detection

and prevention of stroke.

1 INTRODUCTION

Stroke is a devastating illness that affects millions of

people around the world. According to recent

estimates, over 15 million individuals suffer from this

condition every year. While it is true that stroke is a

major health concern in the United States, with one

person experiencing it every four minutes, this is also

a worldwide problem. Stroke is a leading cause of

death and disability on a global scale, with about 6

million individuals dying from it and another 5

million being left permanently disabled. Clearly,

more needs to be done to prevent and treat this

debilitating illness (WSO, 2022).

A medical emergency commonly referred to as a

brain stroke, brain attack, or cerebrovascular accident

(CVA), occurs when the blood supply to the brain is

interrupted. This can be caused by a blockage or

rupture of a blood artery in the brain. When blood

flow is compromised, brain cells start to die due to a

lack of oxygen and nutrients. This can result in

irreversible brain damage, disability, or even death.

There are two primary subtypes of stroke:

hemorrhagic stroke and ischemic stroke. A

hemorrhagic stroke occurs when a blood vessel in the

Corresponding Author

brain bursts and causes bleeding, while an ischemic

stroke occurs when a blood clot blocks an artery in

the brain (A. Kumar, et al., 2023).

High blood pressure, smoking, diabetes, high

cholesterol, obesity, and a family history of stroke or

heart disease are some variables that can increase the

risk of stroke. Lifestyle choices such as poor diet,

inactivity, and stress are also stroke risk factors.

Public education and awareness initiatives can help

reduce the incidence of stroke by promoting healthy

lifestyle choices and encouraging individuals to seek

early medical assistance if they experience stroke

symptoms (R. R. Bailey, 2016). Figure 1 shows the

different type of brain strokes.

Figure 1: Brain Stroke Types – The diagram above

addresses the different types of strokes that can affect a

human brain.

Narayanan, V., Reddy, A., Venkatesh, V., Tutun, S., Norouzzadeh, P., Snir, E., Mahmoud, S. and Rahmani, B.

Brain Stroke Prediction Using Visual Geometry Group Model.

DOI: 10.5220/0012567800003756

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

In Proceedings of the 13th International Conference on Data Science, Technology and Applications (DATA 2024), pages 205-210

ISBN: 978-989-758-707-8; ISSN: 2184-285X

205

Artificial intelligence (AI) has advanced

significantly in recent years, with applications in

everything from autonomous vehicles and medical

diagnostics to voice and image recognition. The

advancement of deep learning, a branch of machine

learning that has made strides in a variety of AI

applications, has been a major force behind this

development. We go over the fundamentals of deep

learning, its uses, and its possible social effects in this

paper.

Deep learning is a vital component of Artificial

Intelligence, which uses artificial neural networks to

model and solve complex classification problems.

These artificial neurons, which function as the

processing and transformation units of these neural

networks, are organized into numerous layers of

interconnected nodes. Deeper layers of neurons learn

to recognize increasingly abstract and complicated

patterns as they learn to recognize and extract

different characteristics of the data.

Deep Learning has the capacity to learn and

improve on its own, without being explicitly

programmed. By processing large amounts of data

and identifying patterns and relationships within that

data, deep learning algorithms can learn to perform

complex tasks such as image and speech recognition,

natural language processing, and even game playing.

In healthcare, deep learning is being used to

develop diagnostic tools to analyse medical images

and data to detect diseases such as cancer and

Alzheimer’s (I. M. Sheikh and M. A. Chachoo, 2022).

These tools have the potential to improve the

accuracy and speed of diagnosis, enabling earlier

detection and better outcomes for patients (L Chen, et

al., 2023; P Bentley, et al., 2014).

Yoon-A Choi et al. proposed a system for

predicting the likelihood of stroke based on real-time

bio-signal data using neural networks. 3,322

electroencephalogram (EEG) and

(electrocardiogram) ECG signals were collected from

stroke patients and healthy individuals. The proposed

model is a Convolutional Neural Network (CNN),

which extracts features from the signals and a long

short-term memory (LSTM) network that models

temporal dependencies. This system gave a model

accuracy of - 93.9 percent and a sensitivity of 96.7

percent (Y-A Choi, et al., 2021). This research

however needed further research to validate the

system’s effectiveness in larger and more diverse

populations.

Vivek S Yedavalli et al. discussed the potential

applications of artificial intelligence (AI) in stroke

imaging, including diagnosis, treatment selection,

and prognosis prediction, using different machine

learning models and neural networks. Models like

Convolutional Neural Networks (CNN), Long Short-

Term Memory (LSTM), Recurrent Neural Networks

(RNN), Support Vector Machines (SVM), and

Random Forest (RF) were trained, and their

accuracies were compared to select the best model. It

was shown that CNN had the highest accuracy at 91

percent. This study used 4 different datasets, namely,

MRI-GENIE, STRIDE, MR CLEAN, and TRACK-

TBI (B B Ozkara, et al., 2023).

Hilbert et al. proposed a deep learning model for

predicting the outcome of endovascular treatment in

patients with acute ischemic stroke. The dataset was

a collection of 92 patients who underwent

endovascular treatment for acute ischemic stroke and

were divided into a training set of 60 patients and a

test set of 32 patients. To improve its performance on

the outcome prediction task, the deep learning model

was trained on a small subset of the training set

(n=10) utilizing transfer learning and fine-tuning

approaches. The model was then assessed using the

test set and the remaining training data. This proposed

system used the VGG-16 model. The model was then

fine-tuned on the training data using a transfer

learning approach, which involves using the pre-

trained weights of the VGG-16 model and training the

final layers on the task-specific dataset. This gave an

accuracy of 80 percent (A Hilbert, et al., 2019).

Sonavane et al. presented a method for detecting

brain stroke using convolutional neural networks

(CNNs) and deep learning models. This is a novel

approach that combines CNNs with deep learning

models to automatically see brain strokes from

computed tomography (CT) images. The dataset is a

collection of 250 CT images, including 150 healthy

and 100 stroke images, which were used to train and

evaluate their deep-learning models. The

methodology tested four different deep learning

models, including a CNN, a deep belief network, a

stacked autoencoder, and a convolutional

autoencoder, to determine the most effective model

for detecting brain stroke. The CNN model achieved

the highest accuracy at 97.6 percent for detecting

brain stroke, followed by the stacked autoencoder

model with an accuracy of 94.8 percent. The authors

also performed a comparative analysis with existing

methods and found that their proposed method

outperformed existing methods for brain stroke

detection. This proposed system’s future research

could explore the use of larger datasets and more

advanced deep learning models to further improve the

performance of the brain stroke detection system (B

R Gaidhani, et al., 2019).

DATA 2024 - 13th International Conference on Data Science, Technology and Applications

206

Mahadevan et al. compared the performance of

traditional hand-crafted features and convolutional

neural networks (CNNs) for diagnosing stroke from

retinal images. The dataset that was used in this

proposed system contained 450 retinal images,

including 150 healthy, 150 hypertensive, and 150

diabetic images, to train and evaluate their models.

The methodology involved in this method comprised

of testing two different approaches: traditional feature

extraction using hand-crafted features, and deep

learning using a CNN. The results of the study

showed that the CNN achieved significantly better

performance than the traditional feature extraction

method, with an accuracy of 96.5 percent compared

to 87.3 percent for the hand-crafted features. The

authors also compared the two approaches and found

that the CNN had higher sensitivity, specificity, and

F1 score for diagnosing stroke from retinal images (R

S Jeena, et al., 2021).

Amitava Nag and his research group did another

research to predict the brain stroke. They applied

Ada-Boost and other boosting methods to investigate

and classify the information of more than 48000

patients achieving high accuracy (S. Mondal, et al.,

2023). The main difference between the current and

Nag’s project is the type of data. We worked on image

data.

In the other similar project, Sachin and Vishal

Jain classified brain tumours with deep learning

models with 98% accuracy. The applied 5, 10 and 20

cross validation folds to realize high accuracy (S. Jain

and V. Jain, 2023). Our project predicts the brain

stroke with Neural Network models.

Emotion recognition project using VGG-16

model accomplished by Srindhar and his research

group in 2023 (S. Vignesh, et al., 2023). Sunil Kumar

et al, applied random forest and VGG-16 methods to

classify bell pepper leaf disease with LBP features

(M. Bhagat, et al., 2023).

Other researchers applied VGG-16 method to

classify and predict different types of diseases and

syndromes, with different type of dataset, and

accuracy. What makes our project different is that we

predict brain stroke using image data with high

accuracy. For this aim we applied different methods,

which VGG-16 predicted the stroke with highest

accuracy among other methods. We worked on a

large image dataset.

https://www.kaggle.com/datasets/afridirahman/brain-

stroke-ct-image-dataset

2 DATA DESCRIPTION

The dataset used for this project is collected from

Kaggle

, an online community, with millions of

diverse datasets available for analyses. We chose a

collection of medical images, specifically computed

tomography (CT) images, of the brain of individuals

who have experienced a stroke and of individuals

who have not experienced a stroke. There are a total

of 2,501 images in the datasets out of which 1,551

belong to individuals who have not experienced

strokes and the remaining 950 belong to individuals

who have experienced a stroke. This is a binary

classification problem where images belong to two

different classes.

3 METHODOLOGY

3.1 Data Pre-Processing

The final image size of the input dataset for the VGG-

16 (Visual Geometry Group) model is 256x256

pixels. In the next pre-processing step we split our

dataset into training (80 percent), testing (10 percent),

and validation (10 percent) sets. To include more

samples in our dataset we implemented an image data

generator that creates different variations of each

image at each epoch. The variations include random

image rotations, horizontal flips, and shifts.

Additionally, zoom and brightness effects are set in

the range of 0.2 and 0.8. After the pre-processing

steps, the transformed images are used by the VGG-

16 model to predict stroke.

3.2 Baseline Model and Methodology

The VGG-16 model which is a 16-layered deep

image classification convolutional neural network

(CNN) architecture is the baseline model in our

research. It is one specific variation of a

Convolutional Neural Networks (CNN), originally

proposed by Simonyan and Zisserman (K Simonyan

and A Zisserman, 2019). This CNN was chosen in

this study based on its prior success at image

recognition. While many variations exist for CNN,

using a configuration that has been previously

validated enables achieving optimal results. The

VGG-16 model differs from other configuration in

its choice of convolutional layers, pooling layers,

and dense layers. In all, VGG-16, has 16

Brain Stroke Prediction Using Visual Geometry Group Model

207

convolutional layers. These are the layers where

learning occurs. This specification facilitates up to

138 parameters that can be trained. The 16

convolutional layers are divided in 5 groups, each

with a pooling layer at the end. At the end of the

stack of layers there a three dense layers. In all,

VGG-16 has 21 layers. In addition to the predefined

layers in the CNN, VGG-16, the convolution layers

are relatively small, with 3x3 filters with a stride

of 1.

The pre-trained VGG-16 model classifies over

1000 images from different categories. We

incorporated the VGG-16 model into our sequential

model with several flattened, dense and output layers.

The output layer consists of a sigmoid activation

function, which is used for binary classification that

consists of two classes i.e. stroke and non-stroke.

Model training was done on 25 epochs and a batch

size of 32. The final sequential model used a

monitoring metric called early stopping that halts the

training process when there is no further

improvement in learning.

In this proposed system, we use ‘Adam’ optimizer

for our model because the Adam optimizer

dynamically adjusts for each parameter based on the

first and second moments of the gradients, which

increases the efficiency of the model performance and

simultaneously, requires low storage space. We

calculate the loss by the binary cross-entropy metric,

which computes gradients correctly and encourages

classification with high accuracy rate.

The model also makes use of early stopping,

which creates an early stopping call back that

monitors the validation accuracy and stops the

training process if the accuracy does not improve for

a specified number of epochs.

4 RESULTS

During the first epoch, the training accuracy started at

61 percent which increased significantly to 83 percent

at the end of the 25th epoch. While predicting

outcomes on the test dataset, the model was able to

correctly predict over 80 percent of the outcomes.

Since the difference between the training and test

accuracies is not too high, we can say that the model

is not prone to overfitting.

From the classification report, the precision for

people with no brain stroke is 0.90, and for the people

with brain stroke is 0.77. This means that there are

more people not detected with brain stroke and fewer

people with brain stroke.

Figure 2: Brain Stroke Detection – The left image

represents the images with no brain stroke and the right-side

image represents images with the prediction of a brain

stroke.

Figure 2 portrays the difference between a normal

brain and a brain that has stroke. The image on the

left side represents the scans of the patients with no

stroke prediction and the images in the right side are

that of the people who have experienced a brain

stroke.

Figure 3: The AUC for Training Vs Validation Loss.

Figure 3 is a graphical representation of the

performance of the Visual Geometry Group Model

during the training and validation phases. This

measures the model’s ability to correctly classify data

points. It is plotted against the number of epochs of

the training process. The image on the left represents

the model loss and exhibits that the model performed

well without any overfitting until the 5

epoch, as the

training and validation loss are close enough to each

other. The image on the right represents the model

accuracy, which measures the overall performance of

the model.

DATA 2024 - 13th International Conference on Data Science, Technology and Applications

208

Figure 4: Confusion Matrix.

Figure 4 is a confusion matrix, that compares the

actual and predicted values of the model on the given

dataset. The confusion matrix displays the least

number of true positives. The number of true

negatives is more than the number of true positives,

which means that there are more predictions for

people with brain stroke.

Figure 5: Classification Report.

Figure 5 portrays the classification report and

provides a detailed evaluation of the model’s ability

to correctly classify instances as is typically used in

classification tasks where the goal is to predict the

class of the given dataset. The proportion of the

positive predictions that actually had stroke is 0.90

and the proportion of actual positive instances that

were correctly identified by the model is 0.86.

5 CONCLUSION

In conclusion, the VGG-16 model shows promising

results in predicting the occurrence of brain stroke

using medical imaging data. Our research

demonstrates that the VGG-16 model can achieve

high accuracy of 80 percent in predicting the

likelihood of stroke by analysing the images of the

brain without overfitting. However, further research

is needed to improve the accuracy and

generalizability of the VGG-16 model, as well as to

explore its potential for other medical imaging

applications. Additionally, it is important to address

ethical and privacy concerns related to the use of

patient data in developing and deploying AI models

for medical diagnosis. Overall, our research suggests

that the VGG-16 model has significant potential as a

tool for predicting brain stroke using medical imaging

data and underscores the importance of continued

research and development in this area.

6 FUTURE SCOPE

In this project we applied neural networks to predict

the stroke. In the next project we apply another

prediction and image analysis methods to validate our

results with higher accuracy. There is a plan to

develop a platform to upload images and define the

possibility of brain stroke.

Data Availability

The dataset used for this project is collected from

Kaggle. https://www.kaggle.com/datasets/afridirahm

an/brain-stroke-ct-image-dataset.

Conflict Interest Statement

There is no conflict of interest declared by authors.

All authors have reviewed and agreed with

manuscript. We state that the submission is original

paper and is not under review at any other journal.

Funding

No Funding has been applied for this project.

Ethical Approval

All subjects gave their informed consent for inclusion

before they participated in the study.

Consent to Participate

Authors consent to participate in this project and we

know that: the research may not have direct benefit to

us. Our participation is entirely volunteer. There is a

right to withdraw from the project at any time without

any consequences.

Consent to Publish

We give our consent for the publication of exclusive

details, that could be included figures and tables and

details within the manuscript to be published in

Computational Brain & Behavior.

Brain Stroke Prediction Using Visual Geometry Group Model

209

REFERENCES

Global Stroke Fact Sheet, World Stroke Organization

(WSO), 2022.

A. Kumar, A. Unnithan, J. M Das, Parth Mehta.

Hemorrhagic Stroke, StatPearls, 2023

R. R. Bailey, Lifestyle Modification for Secondary Stroke

Prevention, PubMed, 2016

I. M. Sheikh, M. A. Chachoo, An enforced block diagonal

low-rank representation method for the classification of

medical image patterns, International Journal of

Information Technology, 2022.

L Chen, P Bentley, D Rueckert, Fully automatic acute

ischemic lesion segmentation in DWI using

convolutional neural networks. NeuroImage: Clinical.

Retrieved February 27, 2023

P Bentley, J Ganesalingam, AL Carlton, K Mahady, S

Epton, P Rinne, P Sharma, O Halse, A Mehta, P

Rueckert, Prediction of stroke thrombolysis outcome

using CT brain machine learning. Neuroimage Clin.

2014 Mar 30;4:635-40. doi: 10.1016/j.nicl.2014.

02.003. PMID: 24936414; PMCID: PMC4053635.

Y-A Choi, S-J Park, J-A Jun, C-S Pyo, K-H Cho, H-S Lee,

J-H Yu, Deep Learning-Based Stroke Disease

Prediction System Using Real-Time Bio Signals,

Sensors, 2021

B B Ozkara, M Karabacak, O Hamam , R Wang, A Kotha,

N Khalili, M Hoseinyazdi, M M Chen, M Wintermark,

V S Yedavalli, Prediction of Functional Outcome in

Stroke Patients with Proximal Middle Cerebral Artery

Occlusions Using Machine Learning Models, Pubmed,

2023

A Hilbert, LA Ramos, HJA Van Os, et, Data-efficient deep

learning of radiological image data for outcome

prediction after endovascular treatment of patients with

acute ischemic stroke, Computers in Biology and

Medicine, 2019

B R Gaidhani, R R Rajamenakshi, S Sonavane, Brain

Stroke Detection Using Convolutional Neural Network

and Deep Learning Models, Conference: 2019 2nd

International Conference on Intelligent Communication

and Computational Techniques (ICCT), 2019

R S Jeena, G Shiny, A S Kumar, K Mahadevan, A

Comparative analysis of stroke diagnosis from retinal

images using hand-crafted features and CNN, Journal

of Intelligent and Fuzzy Systems 41(3):1-9, 2021

S. Mondal, S. Ghosh, A. Nag, Brain stroke prediction

model based on boosting and stacking ensemble

approach, International Journal of Information

Technology, 2023.

S. Jain, V. Jain, Novel approach to classify brain tumor

based on transfer learning and deep learning,

International Journal of Information Technology, 15,

pages 2031–2038, 2023.

S. Vignesh, M. Savithadevi, M. Sridevi, R. Sridhar, A novel

facial emotion recognition model using segmentation

VGG-19 architecture, International Journal of

Information Technology volume 15, pages1777–1787

(2023).

M. Bhagat, D. Kumar, S. Kumar, Bell pepper leaf disease

classification with LBP and VGG-16 based fused

features and RF classifier, International Journal of

Information Technology volume 15, pages465–475

(2023)

O Ozaltin, O Coskun, O Yeniay, A Subasi, A Deep

Learning Approach for Detecting Stroke from Brain CT

Images Using OzNet. Bioengineering. 2022; 9(12):783.

K Simonyan, A Zisserman, Very deep convolutional

networks for large-scale image recognition. arXiv

preprint arXiv:1409.1556. 2019

DATA 2024 - 13th International Conference on Data Science, Technology and Applications

210