BRAIN STROKE.docxBRAIN STROKE.docxBRAIN STROKE.docx

BRAIN STROKE.docx

1. Print ISSN : 2395-1990 | Online ISSN : 2394-4099
Themed Section: Engineering and Technology
Brain Stroke Identification: A Deep Learning –Based Diagnostic
Model Using Neuroimages.
NAME
1
*
1
Senior Scale Lecturer, Department of Computer Science and Engineering, Government Polytechnic, Channasandra, Kadugodi, Bangalore,
India, 560067
[email protected]
ABSTRACT
Stroke is one of the most prevalent causes of death and disability in the world, yet it is both preventable and treated.
Improving clinical outcomes and lowering the burden of illness are significantly aided by early stroke identification
and prompt treatments. Because machine learning techniques may be used to identify strokes, they have garnered a
lot of attention in recent years. Finding trustworthy techniques, algorithms, and characteristics that support
healthcare providers in making well-informed decisions on stroke prevention and treatment is the goal of this
project. We have created an early stroke detection system that uses brain CT scans, a genetic algorithm, and a
bidirectional long short-term memory (BiLSTM) to identify strokes at an extremely early stage in order to
accomplish this aim. Neural networks based on a genetic method are used to identify the most pertinent elements
for categorization in images. These characteristics are then incorporated into the BiLSTM model. The diagnostic
system's accuracy, precision, recall, F1 score, Receiver Operating Characteristic Curve (ROC), and Area under the
Curve (AUC) were all assessed using cross-validation. The overall efficacy of the system was assessed using each
of these indicators. The accuracy of the suggested diagnostic method was 96.5%. Additionally, we contrasted the
suggested model's performance with that of Random Forests, Naive Bayes, Decision Trees, Support Vector
Machines, and Logistic Regression. The suggested diagnosis approach would enable doctors to make well-informed
decisions regarding stroke.
Keywords: Stroke, feature selection, genetic algorithm, LSTM, BiLSTM, CT images, Deep Convolutional Neural
Network (CNN), Feature Extraction, Classification.
I. INTRODUCTION
A stroke happens when a blood vessel in the brain breaks or the blood supply to the brain is cut off. With more
than 6.2 million fatalities each year, it is the leading cause of mortality globally [1]. Many survivors have
infirmities that significantly lower their quality of life. The severity of stroke and its effects on the individual can be
lessened with the use of preventive measures and prompt intervention. Preventing strokes requires early
identification of those who are at risk [2]. Hemorrhage-related stroke and ischemic stroke are the two forms of
1

2. International Journal of Scientific Research in Science, Engineering and Technology (ijsrset.com)
stroke. Ischemic strokes are the most prevalent form, accounting for around 87% of all strokes [3]. Blood flow to a
particular area of the brain is decreased or stopped when a blood artery that supplies blood to the brain becomes
blocked or constricted. A blood clot, cholesterol buildup, or other material entering the brain from other areas of the
body might be the source of the obstruction [4]. Although hemorrhage-related strokes are less frequent, they are
more dangerous and frequently deadly if left untreated. When a blood artery in the brain ruptures or leaks, blood
seeps into the surrounding brain tissue, resulting in a cerebral hemorrhage. This causes the brain's structures to
enlarge and become under strain. This may harm brain tissue and interfere with regular brain activity [5].
This paper develops a methodical approach to stroke detection using machine learning techniques. The primary
goals of the research are to better understand the causes of strokes and develop trustworthy detection models that
support doctors in making well-informed decisions on stroke treatment and prevention. To create prediction models
that identify people at risk of stroke, the project will examine genetic data, lifestyle data, and electronic medical
records. Feature selection techniques will be used in the study to enhance detection performance and glean
insightful information from the data. Additionally, a machine learning model will assist physicians.
This study's objective is to assess the suggested algorithms' efficacy, lucidity, flexibility, and scalability. To
ascertain their relative performance, the suggested models were contrasted with well-established clinical risk
detection techniques using a variety of real-world datasets. To lessen the effects of stroke on patients and the
healthcare system, this study may offer insightful information for clinical practice and customized therapies. This
work aims to address the issue of stroke incidence by using machine learning techniques to forecast the frequency of
strokes. This work enhances the precision and understandability of stroke detection models through the application
of cutting-edge machine learning techniques. By enabling targeted therapies to avoid this crippling condition in the
future, the findings of this study have the potential to completely transform stroke prevention and enhance treatment
outcomes for those who already have it. The following is a list of this study's main contributions:
1) This work uses an image-based dataset to offer a stroke diagnosis method.
2) Using a genetically tuned CNN, the suggested diagnostic approach retrieves valuable features from the CT
images.
3) Based on retrieved characteristics, LSTM, BiLSTM, and genetically adjusted CNNs were evaluated for stroke
classification.
4) Other ML and DL techniques are also used to compare the performance of the suggested diagnostic system
(GA_BiLSTM).
5) By offering early detection, the suggested approach seeks to assist medical professionals in making well-
informed decisions on stroke prevention and treatment.
II. RELATED WORKS
Models for detecting stroke risk that are extremely accurate have been created using machine learning techniques.
These algorithms forecast the risk of stroke in hypertensive patients based on past data from electronic medical
7

3. International Journal of Scientific Research in Science, Engineering and Technology (ijsrset.com)
records [6]. To get the best results in predicting the risk of stroke, a variety of machine learning methods were
employed, including Extreme Gradient Boosting (XGBoost) [7], [8]. Promising outcomes have been shown in the
identification of stroke risk people through the application of machine learning models, particularly ensemble
models that integrate several techniques [9].
Skull stripping methods are covered in detail by the authors of [10], including more recent deep learning-based
methods. This study separates the current methods for skull stripping into two groups: conventional or classical
methods and deep learning or convolutional neural networks. The promise of a few methods is highlighted since
they can be included into standard clinical imaging procedures.
In [11], the authors introduce TDRL, a time-based link prediction model that learns from a growing crime
dataset in the real world using deep reinforcement learning techniques.
The tests demonstrate that compared to conventional machine learning models trained on a single point in time,
the TDRL model trained on a temporal dataset has a higher prediction accuracy.
During the year of the bombing, the TDRL-CNA model correctly predicted the majority of the network
topology's edges, but it was unable to forecast fading edges. Based on the eliminated edges, feature analysis showed
that learning had the least impact on the TDRL-CNA model.
Incomplete datasets make criminal network analysis (CNA) challenging, and the majority of machine learning
approaches depend on supervised learning. The authors looked at developing a model for anticipating hidden links
in criminal networks using deep reinforcement learning (DRL).
The outcomes demonstrate that our method outperforms conventional supervised machine learning approaches
[12].
Using a CNN hybrid structure for artificial neural networks developed by the National Institutes of Health, a
research in [13] looked at clinical brain data CT and predicted multiple stroke scores after 24 hours or a Modified
Rankin Scale score from 0 to 1 over 90 days (or "mRS90"). They were able to detect the mRS90 with 74%
accuracy using this structure.
To categorize brain MRI pictures into normal brains, strokes, and degenerative disorders, a probabilistic neural
network was combined with an integrated wavelet entropy-based spider network graph [14]. As part of their
investigation, the writers methodically looked at illnesses, infectious disorders, and brain tumors. The initial phase
involved processing two-dimensional (2D) brain pictures using a discrete wavelet transform. Spider web
visualizations were used to extract features, and a probabilistic neural network was used to classify them. They
claimed to have a 100% categorization accuracy rate.
Through the analysis of several elements in electronic medical records, the authors of [15] and [16] hope to
enhance the diagnosis and treatment of strokes. They employ principal component analysis and statistics to more
accurately forecast strokes. Patients with heart disease, high blood pressure, average blood sugar, and advanced age
are most likely to have strokes.
In addition, the study recommends utilizing a perceptron neural network with these four characteristics in order
to outperform other benchmarking techniques in terms of accuracy and error rate. Furthermore, the authors offer
7

4. International Journal of Scientific Research in Science, Engineering and Technology (ijsrset.com)
findings using a balanced dataset produced by sub-sampling techniques in order to address the issue of imbalanced
datasets.
It has been demonstrated that the deep learning architecture put out in [17] is extremely accurate in identifying
various intracranial hemorrhage (ICH) subtypes in head CT scans. The system's average accuracy for the three
kinds of hemorrhages—intraparenchymal, subdural, and epidural—is 96.21%. The false positive rate is much lower
than in previous research. Furthermore, a quantitative scoring mechanism included into the device automatically
gauges the amount and thickness of hemorrhagic lesions. This makes clinically meaningful quantification possible,
which is crucial for deciding on an urgent surgical course of therapy.
Grayscale co-occurrence matrix texture data and transformed CT image characteristics are used in a hybrid
feature selection method in [18]. GLCMs, discrete cosine curves, and discrete wavelet transformations are used to
extract features. For classification, the machine learning algorithms Random Tree, Random Forest, and REPTree
are employed; Random Forest achieves the maximum accuracy of 87.97% when combined with GLCM features and
discrete wavelet transformations.
A machine learning-based method for classifying CT images with intracranial hemorrhages is shown in [19].
The Gray Level Co-Occurrence Matrix (GLCM), Discrete Wavelet Techniques (DWT), and Discrete Cosine
Techniques (DCT) are a group of shared characteristics.
The oversampling issue is resolved using SMOTE, and feature subsets are obtained by sequential forward
feature selection. Recall, precision, and a confusion matrix are used to assess classification accuracy. When used in
conjunction with the suggested feature extraction technique, Random Forest produced the greatest accuracy of 87.22
percent.
In [20], the authors examine the results of several preprocessing and model creation techniques and assess how
well a deep learning model performs in identifying cerebral bleeding on CT head images. By using preprocessing
methods and a CNN-RNN architecture, the model's performance was much enhanced and its promise as a tool for
radiologist decision assistance was shown.
We discovered that current studies can categorize strokes with less accuracy based on the literature evaluation
that was done.
Furthermore, the suggested techniques required a lot of calculation. However, we suggest a diagnostic method
that analyzes brain pictures from a CT scan and predicts strokes using a genetic algorithm with bidirectional long
short-term memory (BiLSTM). The previously suggested models' issues with reduced accuracy and increased
computing complexity are resolved by the suggested approach.
III. DESIGN AND IMPLEMENTATION
7

5. International Journal of Scientific Research in Science, Engineering and Technology (ijsrset.com)
Each of the various essential phases that make up the framework for image-based stroke classification is vital to
precise diagnosis and detection. Below is a breakdown of these phases:
Figure-1. System architecture of proposed approach
1. Image Acquisition
The first step is image acquisition, in which pictures are either taken from pre-existing data sources or through a
lens-based system, such medical imaging equipment. The obtained pictures must be sharp, detailed, and distortion-
free regardless of the source. The accuracy of the analysis is greatly increased by high-quality photos, which
guarantee accurate abnormality detection. For the data to remain transparent and reliable, a well-taken picture is
essential.
2. Image Pre-Processing
The image goes through a pre-processing step to standardize its size and quality before examination. The assessment
procedure may become inconsistent due to changes in noise levels, illumination, and picture quality. Therefore, it is
crucial to improve important characteristics, eliminate undesired distortions, and resize photographs to a consistent
standard. This stage guarantees that the classifier is given data that is tuned for precise stroke identification.
7

6. International Journal of Scientific Research in Science, Engineering and Technology (ijsrset.com)
3. Management and Storage of Data
An orderly data storage system is required to enable efficient testing and training. The framework needs well
prepared datasets for training machine learning models because it uses supervised learning. These datasets, which
provide the basis for categorization, are produced from photos gathered during the acquisition stage. To enhance
model performance, photos must be appropriately labeled according to the degree of the stroke.
4. Stroke Detection Classification System
A number of units make up the categorization system, which is intended to improve the visual data and uncover
significant patterns. Based on visual analysis, the classifier, which is in the end, calculates the likelihood of a stroke.
There are two main parts to the classifying process:
a) Unit for Image Preparation
In addition to improving picture quality, this machine makes sure that the important characteristics are separated out
for precise diagnosis. It is made up of many subunits:
To provide a clear portrayal of the scanned region, the noise reduction unit eliminates extraneous colors and artifacts
from the picture.
Image Enhancement and Segmentation Unit: Modifies sharpness, contrast, and brightness to improve the impacted
area. By separating the aberrant area from the typical scanned picture, segmentation makes it easier to extract
features precisely.
b) Component for Feature Extraction
A critical stage in picture classification is feature extraction, which finds distinctive traits important for stroke
detection. This procedure extracts important characteristics like:
Asymmetry
Patterns on the edges
Variations in color
Distance in space
The abnormality's progression
These collected features serve as the classifier's input characteristics and offer crucial details for distinguishing
between different kinds of strokes.
7

7. International Journal of Scientific Research in Science, Engineering and Technology (ijsrset.com)
5. Classification and Identification of Strokes
Determining whether the observed stroke is moderate or severe is the last step. This procedure consists of:
Features of the Input Processing: Every characteristic that has been retrieved is organized and processed for
categorization.
Classifier Engine: To classify photos into preset stroke types, the classifier uses machine learning techniques. For
increased accuracy, sophisticated deep learning methods like convolutional neural networks (CNNs) may be
applied.
The framework facilitates early diagnosis and treatment planning by combining these elements to provide an
effective and dependable approach for stroke detection and classification.
IV.RESULTS AND DISCUSSION
This section outlines the baseline models that are contrasted with the suggested model and provides a summary of
the findings. To assess the suggested model's performance, we contrast it with several Deep Learning (DL)
techniques and Machine Learning (ML) classifiers. All algorithms were employed in this study with their default
settings.
TABLE 1. Performance of DL classifiers.
Our method will lessen the burden on individuals who regularly work in this industry by helping to determine
whether a stroke is positive or negative. This application may be expanded to include the ability to predict several
illnesses. This tactic comprises using the most suitable and accurate AI classification techniques along with the
updated dataset to analyze the prediction approach. CNN stands for Convolutional Neural Network. The goal of this
work is to classify strokes and forecast them using a collection of CT (computed tomography) images. A positive or
negative stroke is used to symbolize the results.
The following actions must be taken in order to put the suggested technique into practice. Install the necessary
packages, including Matplotlib, Tensorflow, Keras Pandas, OpenCV, NumPy, and PIL. The study's foundation is
7

8. International Journal of Scientific Research in Science, Engineering and Technology (ijsrset.com)
real data that was obtained from Hyderabad, India's Virinchi Hospital. In order to train our model, load the dataset.
Prepare the data: Preparation The photos were resized, grayscaled, and reshaped into the proper format. Our model
makes use of this pre-processed training dataset. The dataset is divided into 75% training photos and 25% test
images. Train the CNN models with the selected layers using the training set. Each architecture's training accuracy
and model loss following the predetermined number of epochs (30) following training, the model was tested, and
the outcome is displayed in Fig. 8. Training precision and loss, as well as validation precision and loss, are shown.
Table I. The submitted image's positivity or negativity is predicted by a trained algorithm.
V. CONCLUSION
In this study, we used cutting-edge machine learning techniques to create a reliable and effective model for the
identification and categorization of brain strokes. The model may be deployed in actual healthcare settings because
it has been developed in an approachable environment using Flask and Python. With an astounding predicted
accuracy of almost 90%, the suggested approach proves to be dependable in helping medical personnel diagnose
strokes.
Our model's efficacy is ascribed to its capacity to accurately categorize different types of strokes, extract important
information, and analyze medical pictures rapidly. The system outperforms conventional classification techniques
by including machine learning methods like Support Vector Machine (SVM), Random Forest, and Decision Tree.
With the best accuracy, precision, and recall scores among them, SVM was the most successful classifier. This
model's potential use in hospital settings for early stroke prediction is a major benefit as it may result in quicker
diagnosis, prompt medical intervention, and better patient outcomes. In the treatment of stroke, early diagnosis is
essential, and radiologists and neurologists may find this model to be a useful decision-support tool.
VI. REFERENCES
[1] J. Greene and I. Bone, Understanding Neurology: A Problem-Oriented Approach. CRC Press, 2007.
[2] S. Pouyanfar et al., "A Survey on Deep Learning: Algorithms, Techniques, and Applications," ACM Computing Surveys (CSUR), vol.
51, no. 5, p. 92, 2018.
[3] V. Badrinarayanan, A. Kendall, and R. Cipolla, "SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image
Segmentation," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 12, pp. 2481–2495, 2017.
[4] X. Liu, Z. Deng, and Y. Yang, "Recent Progress in Semantic Image Segmentation," Artificial Intelligence Review, vol. 52, no. 2, pp.
1089–1106, 2019.
[5] A. Gautam and B. Raman, "Towards Effective Classification of Brain Hemorrhagic and Ischemic Stroke Using CNN," Biomedical
Signal Processing and Control, vol. 63, p. 102178, 2021.
[6] M. S. Sirsat, E. Fermé, and J. Câmara, "Machine Learning for Brain Stroke: A Review," Journal of Stroke and Cerebrovascular
Diseases, vol. 29, no. 10, p. 105162, 2020.
[7] T. Durduran et al., "Diffuse Optical Measurement of Blood Flow, Blood Oxygenation, and Metabolism in a Human Brain During
Sensorimotor Cortex Activation," Optics Letters, vol. 29, no. 15, pp. 1766–1768, 2004.
7

9. International Journal of Scientific Research in Science, Engineering and Technology (ijsrset.com)
[8] F. Ouhmich et al., "Liver Tissue Segmentation in Multiphase CT Scans Using Cascaded Convolutional Neural Networks," International
Journal of Computer Assisted Radiology and Surgery, vol. 14, no. 8, pp. 1275–1284, 2019.
[9] G. E. Hinton, S. Osindero, and Y. W. Teh, "A Fast Learning Algorithm for Deep Belief Nets," Neural Computation, vol. 18, no. 7, pp.
1527–1554, 2006.
[10] M. Toğaçar, Z. Cömert, and B. Ergen, "Classification of Brain MRI Using Hyper Column Technique with Convolution Neural
Network and Feature Selection Method," Expert Systems with Applications, vol. 149, p. 113274, 2020.
[11] S. Kuraparthi et al., "Brain Tumor Classification of MRI Images Using Deep Convolutional Neural Network," Traitement du Signal,
vol. 38, pp. 1171–1179, 2021.
[12] K. R. Madhavi, P. Kora, L. V. Reddy et al., "Cardiac Arrhythmia Detection Using Dual-Tree Wavelet Transform and Convolutional
Neural Network," Soft Computing, vol. 26, pp. 3561–3571, 2022. [Online]. Available: https://doi.org/10.1007/s00500-021-06653-w
[13] V. S. Kumar, K. G. Suma, and M. Sakthivel, "Segmentation of Hyperspectral Satellite Image Based on Classical Clustering Method,"
International Journal of Pure and Applied Mathematics, vol. 18, no. 9, pp. 813–820, 2018.
[14] P. Yadlapalli et al., "Breast Thermograms Asymmetry Analysis Using Gabor Filters," E3S Web of Conferences, vol. 309, 2021.
[Online]. Available: DOI:10.1051/e3sconf/202130901109.
[15] P. Karuppusamy, "Hybrid Manta Ray Foraging Optimization for Novel Brain Tumor Detection," Journal of Soft Computing Paradigm
(JSCP), vol. 2, no. 3, pp. 175–185, 2020.
[16] T. Vijayakumar, "Classification of Brain Cancer Type Using Machine Learning," Journal of Artificial Intelligence, vol. 1, no. 2, pp.
105–113, 2019..
7