0% found this document useful (0 votes)
34 views23 pages

Version 1

Uploaded by

iqra
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
34 views23 pages

Version 1

Uploaded by

iqra
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 23

BRAIN HEMORRHAGE DETECTION USING DEEP LEARNING,MACHINE

LEARNING AND HYBRID MODELS

Version#1

BY

ALIA SALEEM(MSCS06)

IQRA MUSHTAQ(MSCS08)

A review paper submitted to the Department of Computer Science in partial fulfillment of the
requirements for the degree of M.S.C.S in Computer Science.

Department of Computer Science and IT January 2023


Abstract:

The brain is considered as the sensory system's central organ. It is regarded as the human body's control
unit because it directs the majority of its activities. Unlike all other body parts, the brain suffers from
some disorders when it is damaged or harmed. Every day, many people with brain injuries are admitted to
emergency rooms. As a result of the large number of slices analyzed by the doctor for each patient and to
accelerate the diagnosis the development of a precise computer aided diagnosis systems becomes much
recommended. Given the gravity of these issues, researchers have developed a variety of brain
hemorrhage detection techniques. Brain hemorrhage is detected using lesion parameters such as
symmetry, color, size, shape, and so on. This paper provides an in-depth systematic review of deep
learning, machine learning and hybrid techniques for brain hemorrhage detection and also provides the
comparison of these algorithms. So, the study concluded that from the relevant analysis that the results
obtained from hybrid features extracted from the ANN models are best with accuracy 99.3%, precision of
99.36%, sensitivity of 99.5%, specificity of 99.57% and AUC of 99.84% as compared with other models.
The study looked at research papers from reputable journals that were relevant to the topic of brain
hemorrhage diagnosis. To aid comprehension, research findings are presented in the form of tools, graphs,
tables, techniques, and frameworks.

Keywords: Deep learning, machine learning, brain hemorrhage, Random Forest, CAD systems, neural
networks, hybrid models, cerebral

1. Introduction

In the last few decades, machine learning algorithms have been extensively used to analyze medical
datasets [1] [2] [3]. Nowadays, most modern hospitals are well equipped with data monitoring and data-
collecting devices. The combinations of rich, diverse data obtained from these modern devices and
powerful machine learning algorithms are principles for developing computer-aided detection (CAD)
systems. Among a huge amount of machine learning algorithms, probabilistic models, support vector
machines (SVMs) and artificial neural networks (ANNs) have been commonly used in CAD systems [2]
[3]. Especially, there has been a great deal of significant research that uses ANNs in medical decision
support. In recent years, deep learning exploiting the super strength of high performance computing
(HPC) systems have been becoming an interesting approach to image processing. Particularly, deep
neural networks (DNNs) have won numerous contests in pattern recognition and machine learning
[6].DNNs need to be supported by HPC systems to become a powerful and useful tool. Due to the strict
requirement of hardware and expensive medical datasets, DNNs have been rarely used in medicine
generally and CAD systems particularly. Since 2015, there have been a few remarkable publications
applying DNNs to medicine.

The brain is considered to be the central organ of the sensory system. It is considered as the control unit
of the human body as it directs most of the activities of a human body. Unlike all the body parts of a
human body, the brain also faces some disorders when our brain is damaged or harmed .Hemorrhage [1],
[2] is a medical term referring to bleeding within or out of the body. Internal bleeding of the brain is
known as Brain Hemorrhage [3].There is three main causes of brain hemorrhage are blood pressure,
consumption of alcohol, and heredity. A patient’s response to a brain hemorrhage depends on the size of
the hemorrhage and the amount of distension.

The hemorrhage is divided into categories using the CT-images. Aforementioned as


IntraventricularHemorrhage(IVH),IntracerebralHemorrhage (ICH), Subarachnoid hemorrhage (SAH), Epidural
hemorrhage (EDH), and Subdural .According to World Health Organization (WHO), stroke is "rapidly developing
clinical signs of focal (or global)disturbance of cerebral function, with symptoms lasting 24 hours or longer or
leading to death, with no apparent cause other than of vascular origin”. In a WHO report report of 2009 patients
suffer from stroke was 15 million, in which 5 million died and 5 million were disabled [16]. To diagnose the
patient’s disease through the medical image, doctors usually suggest MRI or CT scan. CT scans are widely used
compared to brain MRI. A Computed Tomography scan or CT scan is a medium used to detect brain abnormalities.
It is a diagnosis where images of inside organs of the human body are taken using a series of X-rays. CT scanned
images are more convenient than X-ray images because, in regular x-rays, there is only one tube fixed which can
send rays in only one direction, but CT scan uses motorized x-rays which keep rolling over the human body, which
helps to get a clearer image of the body. For our research, we have considered using CT brain images for detecting
and classifying brain hemorrhage. The identification of the brain hemorrhage is a challenging phase because it is
caused by internal bleeding in the head [8] and the medical experts also need years of experience to identify the
region of bleeding in the CT scan. There is a life of the patients is at risk and every single life is important. In
previous studies, the researchers put their efforts and knowledge to make the perfect system to diagnose the exact
region of brain hemorrhage but failed because of deficiencies like diagnosis processes took too much time and
performance evaluations illustrates not enough results to save the life of every single patient. With the help of
Computer-Aided Detection (CAD), it will be easier for radiologists to analyze data and generate reports. For the
help of the radiologists on their task of interpreting medical images, Computer Aided Detection has been improved a
lot in the last two decade

(Monessen teal., 2012) [4]. Machine learning can be used for analyzing and detecting brain hemorrhage
from CT scan data.CNN,GAN, RF are the common methods that are used to detect brain hemorrhage
from image data. Regarding to CAD systems in terms of brain hemorrhage, in this study, we propose an
approach base on deep and machine learning algorithms to distinguishing the status of brain-hemorrhage-
suspected cases.Basically,1st prepossess the input images. The preprocessing is the phase of filtering,
reshaping and enhance the data set quality to increases the performance of the deep learning models.
There are different methods used in this study like re-sizing, flipping and augmentation of images are
applied as a preprocessing to enhance the quality and quantity of the images data. Then models are
applied on this preprocessed data. After the feature extraction we RF ML-algorithm used for testing.
Figure 1:Process of Brain hemorrhage detection

Many researchers proposed different techniques to increase the efficiency, accuracy and performance of
the detection process. In this paper, different methods by using various models in single and in hybrid
style used to detect and diagnose hemorrhage discussed in detail using systematic literature review. The
main goal of this paper is to present an exhaustive systematic literature review of the methods of Brain
hemorrhage detection. Over the last decade, different methods of detection using CNN models have been
developed, such as Google Net, Alex Net, Res Net-xx, Mobile Net and VGG are proposed. It is important
to classify, analyze and summarizes the proposed methods and models. Taxonomy of this review paper
consists of total seven sections including introduction, the remainder of the paper is organized as follows.
Section 2reviews the previous relevant studies. Section 3 describes the methodological mechanism of the
proposed methods for analyzing and evaluating CT images for early detection of the hemorrhage disease.
Section 4 summarizes the experimental results of each proposed method. Section 5 presents a discussion
of the proposed methods and compares the diagnostic results of all the proposed methods. Section 6
concludes the study,

Related Work this section presents a set of related studies to the diagnosis of hemorrhage data set. This
study was distinguished by the diversity of methods and methodologies for analyzing the data set and
reaching superior results in indicating which image contained a hemorrhage. In already existing method
was to diagnose the hemorrhage data set using pre-trained CNN models, namely Google Net, ResNet-50
and Alex Net, which achieved accuracies of 94%, 91.7% and 91.5%, respectively. The second purposed
Method is Machine learning Models and the third proposed method which is resects advancement in
technology is hybrid technique between deep learning (Google Net, ResNet-50 and Alex Net).
Sr Model Title Contribution Title Type yea Re
no r f.
1. Alex Net model Awwal et al. presented an Alex Net model and Application of Deep Learning in Journal 2019 [3]
and Alex Net- Alex Net-SVM for detecting Cerebral Neuroradiology : Brain
SVM hemorrhages through CT for the hemorrhage Hemorrhage Classification Using
data set. The experiments proved the Transfer Learning
superiority of Alex Net-SVM over the rest of
the models.
2. CNN model Anas et al. proposed a CNN model for Asian Journal of Medical Journal 2022 [4]
Diagnosing 200 samples from the data set with Technology (AJMedTech)
secondary data for the classification of
hemorrhages. This method achieved an
accuracy of 93.14%
3. SVM and CNN Sofia et al. proposed a CNN and supervised Analysis of Intracranial Journal 2021 [5]
models machine learning to diagnose the data set as Hemorrhage in Ct Brain Images
bleeding or healthy. Noise and edge extraction Using Machine Learning and
and classification of data set done by using Deep Learning Algorithm
CNN and SVM
4. Bayesian-based Xiaohong et al. proposed a method based on Deep Learning-Based Detection Journal 2021 [6]
deep learning deep learning to classify and segment the and Diagnosis of Subarachnoid of
and hybrid cerebral hemorrhage zone. Bayesian-based Hemorrhage Healthca
models deep learning and hybrid models were thus re
applied to classify the bleeding site effectively. Engineer
ing
5. ResNet-50 Tomasz et al. proposed the ResNet-50 model to Intracranial Hemorrhage Asian 2022 [7]
determine bleeding quickly and efficiently. The Detection in Ct Scan Using Deep Journal
system achieved an accuracy of 93.3% and Learning Of
recall of 76% Medical
Technol
ogy
6. CNN-DS model Ali et al. proposed a deep learning methodology A fast and fully-automated deep- Journal 2020 [8]
with deep supervision (CNN-DS) to segment the learning approach for accurate
lesion area and measure the volume of bleeding hemorrhage segmentation and
within the brain. volume quantification in non-
contrast whole-head CT

7. d GC-SDL Synergic deep learning extracted deep feature Synergic deep learning model– Article 2022 [9]
maps and categorized them with softmax, based automated detection and
achieving an accuracy and sensitivity of 95.73% classification of brain intracranial
and 94.01%, respectively hemorrhage images
8. Google Net diagnose the hemorrhage dateset using pre- Multi-Method Diagnosis of CT Journal 2022 [1]
trained CNN models, namely Google Net, which Images for Rapid Detection of
achieved accuracies of 94%. Intracranial Hemorrhages Based
on Deep and Hybrid Learning
9. ResNet-50 diagnose the hemorrhage data set using pre- Multi-Method Diagnosis of CT Journal 2022 [8]
trained CNN models, , which achieved Images for Rapid Detection of
accuracies of 91.7% Intracranial Hemorrhages Based
on Deep and Hybrid Learning
10. Alex Net diagnose the hemorrhage dateset using pre- Multi-Method Diagnosis of CT Journal 2022 [8]
trained CNN models, , which achieved Images for Rapid Detection of
accuracies of 91.5% Intracranial Hemorrhages Based
on Deep and Hybrid Learning
2. Research framework

The goal of this systematic literature review was to identify and classify the best available approaches to
brain hemorrhage detection using neural networks (NNs).Systematic reviews of the literature collect and
analyze existing studies using predefined evaluation criteria. Such reviews aid in determining what is
already known in the relevant field of study.The current systematic literature review included research
papers based on deep neural network (DNN) techniques that were relevant to brain hemorrhage
detection.The first step in this systematic review was to define the review framework. In the systematic
literature review, it consisted of an overall plan that was followed. The plan was divided into three layers:
planning, data selection and evaluation, and results generation and conclusion.

Search strategy

A systematic and well-planned search is critical for gathering useful material from the desired domain's
searched data. A thorough search was conducted in this step to extract meaningful and relevant
information from the mass of data. We developed an automated search mechanism for filtering data from
all sources into the desired domain.

Research sources

To extract information relevant to NN techniques for brain hemorrhage detection, we conducted an initial
search on reputable search engines such as IEEE Xplore, ACM, Springer, Elsevier, and Google Scholar.
The primary search filtered out basic research material related to the underlying topic. The selected
research papers and conference proceedings were then examined further in light of the evaluation criteria.
Methodology

SCIENTIFIC DATABASES

• SPRINGER
• IEEE
• ACM
• GOOGLE SCOLER
• MDPI

TOTAL PAPERS=1000

300
bas SEARCH ON BASES OF
es KEYWORDS OF PAPER
Mac TITLE
hin
e Le 75
SEARCH ON OF PAPER
bas
TITLE es
Mac
hin
25 25 e Le
SEARCH ON BASES SEARCH ON BASES SEARCH ON BASES 25
DEEP LEARNING 225
MACHINE LEARNING HYBRID MODELS
25b Bas
ases es
5 5
Mac 5 Mac
GOOGLE NET hin SVM ALEXNET + SVM hin
e Le e Le
GOOGLENET + (GLCM
AND LBP)
5 5 5
ResNet-50 + (GLCM and
RESNET-50 MLP GOOGLENET + SVM
LBP) ANN AlexNet +
(GLCM and LBP)
RESNET-50 + SVM

AlexNet.
5 5 5
ALEXNET. RF ANN ALEXNET + (GLCM
AND LBP)
3.1 Deep learning techniques for brain Hemorrhage Detection

Deep Learning (DL) techniques are used to improve medical diagnosis platforms. This technology is
making the diagnosis practice of brain issues easier for medical practitioners to analyze and identify
diseases with an assured degree of precision and performance. In our study, neural networks were trained
to classify images and to distinguish between hemorrhages. We searched for different techniques of
learning, such as, AlexNet, Google Net and ResNeT-50 for brain hemorrhage detection systems. Research
related to each of these deep neural networks is discussed in detail in this section.

3.2 Google-Net for brain hemorrhage detection

The 2014 ILSVRC competition was won by Google Net (or Inception Network).It was successful in
achieving an error rate of 6.7% in the top 5, which was approximately equivalent to human performance.
Google created the model, which has 22 layers and 5 million parameters, as well as a smarter
implementation of the classic LeNet architecture. This is based on the Inception module's concept. The
core concept underlying Inception modules is that rather than creating convolutional layers with distinct
hyper parameters in individual layers, we execute the entire convolution simultaneously to generate a
result comprising matrices of all the filter operations combined. The Inception module is a neural network
architecture that uses convolutions with different filters to detect features at different scales and
minimizes the computational cost of training a large network through dimensional reduction.

Figure 2 Architecture of Google Net


In this model the 1st layer is convo which uses the filters of size
7x7.the primary purpose of this layer is to reduce the input image but
not lose the spatial information. The second convo layer has a depth
of two and leverages the 1x1 convo blocks, which as the effect of
dimensionality reduction.Dimensionalityreduction through 1x1 convo
blocks allows the decrease of computational load by lessening the
layers' number of operations. These max-pooling layers' purpose is to
down sample the input as it travels through the network. This is
accomplished by reducing the height and width of the input data.
Figure 3: Partition of Google Net. GoogLeNetsupported by
TensorFlow library were pre-trained with the Image Net database and have impressive power to
discriminate several kinds of real-world imagery.

3.3 ResNet-50 for brain hemorrhage detection

After AlexNet, the most revolutionary development in the field of CNN architecture development occurred with
ResNet-50 or Residual Networks.This model has a significant disadvantage — the 'Vanishing Gradient Problem'.
Because the gradient value decreases significantly during backpropagation, so,there is little change in weights.
ResNet is used to overcome this. It employs the "SKIP CONNECTION" technique. The "identity shortcut
connection" was a revolutionary idea integrated into this architecture, which involves transferring the results of a
few layers to some deeper layers, thus skipping certain layers in between, allowing the network to train effectively
over thousands of layers without degrading long-term performance. A deep 1001-layer ResNet-1001 finished in the
top 5 of the 2015 ILSVRC Contest with a 3.57% error rate (beating human-level performance on the data set),
winning all of the challenge's prizes in categorization, sensing, and localization.ResNet's fundamental breakthrough
was that it enabled us to train immensely deep neural networks with 150+ layers. It is an innovative neural network
that was first invented by Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun in their 2015 computer vision
research paper titled 'Deep Residual Learning for Image Recognition'.

Now we’ll talk about the architecture of ResNet50. The architecture of ResNet50 has 4 stages as shown in the
diagram below. The network can take the input image having height, width as multiples of 32 and 3 as channel
width.Stage 1 of the network starts and it has 3 Residual blocks containing 3 layers each. The size of kernels used to
perform the convolution operation in all 3 layers of the block of stage 1.As we progress from one stage to another,
the channel width is doubled and the size of the input is reduced to half. In this particular work, the ResNet-50 was
quite well, and transfer learning was used to detect brain hemorrhage in images (CT scan). The additional layers are
replacing the fully connected layer of the ResNet-50. These additional layers are – a dropout layer with a dropout
probability of 50% – this drops 50% of the parameters randomly and lessens overfitting, followed by a Rectified
Linear Unit (ReLU) activation layer with L2 regularization regression. The value of the regularization parameter is
0.02. Regularization is needed to prevent over fitting of the model, and it is done so by adding a penalty with the
cost function. Next is another dropout layer with a 50% dropout probability. And finally a softmax activation layer
with L2 regularization and an output layer with a size of 3for the 3 classes of data.

Brain hemorrhage

NonBrainhemorrhage

Figure 4: Brain hemorrhage detection using ResNet-50


3.4 AlexNet for brain hemorrhage detection

Alex Net is the first convolutional neural network to win the Image Net Large Scale Visual Recognition Contest
(ILSVRC) in 2012 [10]. This deep structure is made up of eight main layers, the first five of which are mostly
convolutions and the last three of which are fully connected layers. An activation function layer follows each
convolution layer. In AlexNet, a max pooling is used after each convolution layer to reduce network size.
Furthermore, after the first two fully connected layers, a dropout layer is added to help reduce the number of neurons
and prevent over fitting . Finally, a layer is added after the previous one to classify the input data.A pretrained model
(AlexNet) is used, and the final fully connected layer (FC8) is disconnected to add a new layer with two output
neurons corresponding to the two CT brain image categories. It should be noted that the weights of this layer are
chosen at random. The remaining five convolutional layers, on the other hand, Figure 5:AlexNet to detect
brainhemorrhage are kept in the network to share the learned parameters, specifically weights. These weights have
already been trained on large datasets such as ImageNet to extract high-level features from the input data.When
transferring AlexNet knowledge to a haemorrhage classification task, these weights can act as a powerful extractor
of various levels of abstraction from input data features.

Ref Training algorithm Discription Results Limitations

AlexNet Used for any object- Accuracy=91.5% The low training


detection task. effectiveness in the
[9] Precision =93.3% face of “exploding”
and “vanishing”
Sensitivity =91.4%
gradients.

ResNet-50 Used for excellent Accuracy=91.7% If the network is


generalization too shallow, the
[5][10] performance with Precision=91.1% learning might be
fewer error rates very inefficient.
Sensitivity =92.3%
on recognition
tasks

GoogleNet Used for face Accuracy=94% More parameters


detection and also means that your
[12][8][11] recognition, Precision=93.5% model is more prone
adversarial training to overfit.
Sensitivity=93.4%

4 .Machine learning models for brain hemorrhage detection

Machine learning (ML) has enormous potential for transforming healthcare. While published studies have
demonstrated the value of ML models in interpreting medical imaging exams, these are frequently tested
in laboratory settings. Case studies that have documented successes and failures in the translation of these
models into clinical environments best demonstrate the importance of real-world evaluation.
Demonstrating generalizable ML model performance under real-world conditions is a critical prerequisite
for clinical adoption of these technologies.so in this study we discussed ML models like SVM,RF and
VG-19 that are used to distinguish between images weather it is hemorrhagic or non-hemorrhagic.
4.1 SVM (Support Vector Machine)

The Support Vector Machine (SVM) is a machine learning algorithm used for binary classification. SVMs work by
determining a hyperplane that best separates data points from different classes. The kernel used by the SVM defines
the nature of the hyper plane, with a linear kernel being one of the simpler examples. Undoubtedly, SVMs have been
used in similar studies, such as microwave measurements for breast cancer detection and other electrical impedance
frameworks for various tissues. Electrical impedance measurements from a multi-frequency sweep (electrical
impedance spectroscopy, EIS) of various tissues were used to train and test an SVM to classify 'breast' and 'not
breast'.

Figure 6:Brain hemorrhage detection using SVM

The SVM model is built around the creation of a hyper plane that best separates observations from the
two classes. A two-dimensional (2-D) hyper plane (a line) that separates observations classified as +1 or
1. During the training phase, a mathematical model of the hyper plane and margin is created using n-
dimensional training observations (n number of features). The hyper plane is used to determine whether
future observations are in the +1 or 1 class. When the data is not perfectly separable (no margin exists that
guarantees no observations between it and the hyper plane), "soft" margins can be used to ignore outliers.
The kernel, which defines the function used to generate the classifier, is an important parameter when
using SVM classifiers.

4.2 Random Forest (RF)

Random Forest (RF) is a machine learning algorithm that uses the concept of Decision Trees to classify
data (DTs). It is capable of both regression and classification tasks. It is also widely used in the
classification of medical images. The Random Forest algorithm works in the same way that a forest does.
A random forest is made up of independent decision trees that are built with random values. The number
of decision trees determines prediction accuracy. The greater the number of decision trees in the forest,
the greater the accuracy. To impute missing values from a data set, a random forest algorithm is used. It is
also capable of handling a large number of datasets. It is used in classification and regression problems
due to its ease of use and flexibility. It determines the significance of various variables in categorization.
According to a journal article [20], Random Forest showed an average precision of 93.10% in classifying
2400 medical images, which is 4% higher than the MSVM method. The random forest training method
employs the standard bootstrap aggregation approach.
Figure 7: Architecture of RF model

Assume there are a data set D and multiple decision trees, such as DT1, DT2, DT3, and DTN. Different
samples from the data set D, such as S1, S2, S3, SN, will be given to DT1, DT2, DT3, and DTN decision
trees. During the training period, different decision trees will produce varying levels of accuracy.
Following that, the mean of the accuracy will be calculated, which will be the random forest's accuracy
level.

4.3 Multi-layer Perception (MLP)

MLP is an ANN model that belongs to the feed forward network class. In 1958, Frank Rosenblatt created
the perception algorithm at the Cornell Aeronautical Laboratory [22]. It is a supervised learning model
that produces a set of outputs from a variety of inputs. It trains the network using back propagation. This
back propagation method aids in calculating the gradient of a loss function across all network weights.
The multi-layer perceptron (MLP) is used in stock analysis, image recognition, spam detection, and
election vote prediction. This method consists of three layers:

⚫ Input layer

⚫ Hidden layer

⚫ Output layer

The interconnected nodes exchange information by using these layers.The input layer receives an input
that will result in output and routes it through the output layer. This network must have at least one
hidden layer to compute and perform various operations on the input data.

Figure 8: Brain Hemorrhage detection using MLP


Results in the form of table

Ref Training Algorithm Description Results Limitation

It can handle both Accuracy=97.35%. Long training time


classification and for large datasets.
[15] SVM regression on Precision= 96.5%
linear and non-
linear data.

[12],[14] RF It utilizes Accuracy=92.7% The main


ensemble learning, limitation of
which is a Precision=92.5% random forest is
technique that that a large
combines many number of trees
classifiers to can make the
provide solutions algorithm too slow
to complex and ineffective for
problems. real-time
predictions

[13],[14] MLP It is suitable for Accuracy=97.2% Each node is


regression connected to
prediction Precision=92.7% another in a very
problems where a dense web
real-valued resulting in
quantity is redundancy and
predicted given a inefficiency.
set of inputs

4.4 Hybrid of Deep Learning and Machine Learning

In this part, a novel approach for CT image diagnostics is described, namely a hybrid technology
combining the deep learning models used in this work (Google Net, ResNet-50, and Alex Net)
with a machine learning (SVM) algorithm for early and quick identification of hemorrhages.
CNNs demand high specifications and expensive computer resources, and training a dataset takes
a long time, therefore CNN models encounter these issues. As a result, hybrid approaches need
medium-cost computer resources that can train the dataset quickly. As a result, this method
addresses the issues that CNN models confront. Hybrid approaches are made up of two parts:
CNN models and SVM algorithms. The first block (Google Net, ResNet-50 and Alex Net)
receives the enhanced CT images, extracts the deep feature maps, and sends them to the second
block [15]. The SVM technique is used in the second block to classify the feature maps with great
accuracy and efficiency. Figure 9 depicts the approach of the hybrid methods, which include two
blocks termed Google Net + SVM, ResNet-50 + SVM, and Alex Net + SVM for CT image
diagnostic for early identification of hemorrhages.
Figure 8 Hybrid methodology between deep learning and SVM. (a) GoogleNet+SVM. (b) ResNet-50 +
SVM. (c) AlexNet + SVM.

Sage et al.[19] proposed intracranial hemorrhage detection in head ct using double-branch


convolutional neural network, support vector machine, and random forest. The model
trained for each hemorrhage subtype is based on a double-branch convolution neural
network of ResNet-50 architecture. It extracts features from two chromatic
representations of the input data: a concatenation of the image normalized in different
intensity windows and a stack of three consecutive slices creating a 3D spatial context.
The joint feature vector is passed to the classifier to produce the final decision. We tested
two tools: the support vector machine and the random forest. The experiments involved
372,556 images from 11,454 CT series of 9997 patients, with each image annotated with
labels related to the hemorrhage subtypes. We validated deep networks from both
branches of our framework and the model with either of two classifiers under
consideration. The obtained results justify the use of a combination of double-source
features with the random forest classifier. The system outperforms state-of-the-art
methods in terms of F1 score. The highest detection accuracy was obtained in intraventricular
(96.7%) and intraparenchymal hemorrhages (93.3%) [19].

The other combination of features extracted by CNN models (Google Net, ResNet-50, and Alex
Net) with hybrid features derived by GLCM and LBP techniques. This suggested technique is
distinguished by its use of low-cost computer resources and its quickness in training the dataset.
The suggested system's mechanism is as follows. The CT images are then input into the CNN
models, where each model extracts the feature maps and puts them in a feature vector, resulting in
4096 features per picture. Therefore, the dataset is represented in a matrix of features with a size of
7032 × 4096. Due to the high-dimensional features, the features matrix is fed to the PCA
algorithm [16] to reduce the dimensions and choose the essential representative features for each
image. Then, the features matrix becomes 7032 × 1024 in size.

Following recovery, the hemorrhagic region is split and segregated from the healthy area. When
features are extracted from the full picture, features for the damaged and healthy areas are
extracted, resulting in incorrect diagnostic results. This technique is regarded as one of the most
crucial and sophisticated image processing techniques. The chosen region growth algorithm was
employed in this investigation. The algorithm clusters similar pixels together such that the pixels
in each zone are identical. The technique uses an incremental approach, thus the image is divided
into multiple parts, each of which begins with one pixel and increases progressively by gathering
comparable pixels in one region. As a result, the method is repeated until each pixel is allocated to
its own zone. And no comparable pixels can exist in two locations. The method is repeated until
the segmentation is complete and the desired region is acquired. And no comparable pixels can
exist in two locations. The procedure following the fragmentation process, small holes in the areas
of interest that do not correspond to the region of interest are left. As a result, using the
morphological technique is one way to improve the photos after segmentation to generate more
improved images. The technique looks for holes in the area of interest. It handles them using a
variety of procedures based on wrapping the structural element around the picture, such as
opening, closure, erosion, and dilation. The third step is the application of GLCM and LBP
algorithms to extract feature from the lesion area. The GLCM algorithm produces 13 statistically
features [17], while the LBP algorithm produces 203 features [18]. Then, the features of the two
algorithms are combined into one feature vector so that each vector represents 216 features. Thus,
the dataset becomes represented in the feature matrix with a size of 7032 × 216. Fourth, the
resulting features from CNN models after dimensionality reduction are combined with the features
extracted by the GLCM and LBP algorithms and stored in a new feature matrix so that its size
becomes 7032 × 1240. Fifth, the hybrid features between the CNN models and the (GLCM and
LBP) algorithms are fed to an ANN for classifying them with higher accuracy and efficiency than
the other proposed methods. Figure 10 shows the basic structure and methodology of extracting
features using CNN models and GLCM and LBP algorithms, merging them and then feeding them
to the ANN algorithm for classification.

Splitting Dataset: The performance of all suggested networks was assessed using the identical
CT images from the bleeding dataset. The dataset contains 7032 CT scans separated into two categories.
Hemorrhage and non-hemorrhage. During the training, validation, and testing phases, the dataset was
segmented by subject. Table 1 shows the data set breakdown throughout all stages, with 80% for the
training and validation phases (80:20) and 20% for the testing phase. First, 14 subjects (patients) with
cerebral hemorrhages were assigned to the training and validation phase, representing 2152 CT pictures,
and four subjects (patients) with cerebral hemorrhages were assigned to the testing phase, representing
537 CT images[20]. Second, for the healthy images, 22 subjects were assigned to patients who did not
have cerebral hemorrhages, representing 3474 CT images for the training and validation phase, and five
subjects were assigned to patients who did not have cerebral hemorrhages, representing 869 CT images
for the testing phase. The suggested systems in this study were constructed in MATLAB 2018b using
Intel® i5 6th generation computer requirements, 12 GB RAM, and a 4-GB GPU [20].

Figure 9 Basic methodology of hybrid feature extraction technique between CNN models and GLCM and LBP
algorithms and their classification by an ANN

Environment Set-Up

All methods proposed in this study were evaluated with the same measures—accuracy, precision,
sensitivity, specificity, and AUC—for diagnosing the CT images of the hemorrhage dataset. Equations
(1)–(4) illustrate the performance of the proposed systems by Bandai Al-Mekhlafi, Z.G et al [21,22]. It
is noted that the equations contain variables that represent correctly and incorrectly classified images. The
equations’ variables were obtained from the confusion matrix produced by all the proposed systems. A
confusion matrix is a tool for assessing the performance of systems on the hemorrhage dataset. The
confusion matrix contains all samples of the dataset, either correctly classified as TP or TN or incorrectly
classified as FP or FN:

Accuracy = TN + TP TN + TP + FN + FP × t100% (1)

Precision = TP TP + FP × 100% (2)

Sensitivity = TP TP + FN × 100% (3)

Specificity = TN TN + FP × 100% (4)

where true positive (TP) is the number of correctly classified hemorrhage CT images, true negative (TN)
is the number of correctly classified non-hemorrhage CT images, false positive (FP) is the number of CT
images that are not hemorrhages but are classified as hemorrhages, and false negative (FN) is the number
of hemorrhage CT images classified as non-hemorrhage CT images. The AUC is the area under the curve
of the ROC plot, which is defined as the true positive rate against the false positive rate at various
threshold values.

Results of Hybrid Deep Learning with the SVM Algorithm

This section evaluates the performance of hybrid approaches for CT image diagnostics that combine deep
learning networks (GoogLeNet, ResNet-50, and AlexNet) and the SVM algorithm for quick identification
of the haemorrhage dataset. There are several reasons to use this approach, including its necessity for
medium-spec computer resources, quickness in training the dataset, and efficiency in super-diagnostics.
This approach was proposed of two parts: deep learning models and the SVM. Deep learning models
extract deep feature maps and pass them to the SVM algorithm for high-accuracy classification. Table 5
highlights the performance of the hybrid approaches for analysing CT images for quick diagnosis of
cerebral haemorrhages, namely GoogLeNet + SVM, ResNet-50 + SVM, and AlexNet + SVM. The
GoogLeNet + SVM system was shown to be somewhat superior to the other systems. The accuracy of the
GoogLeNet + SVM network was 97.4%, the precision was 97.43%, the sensitivity was 97.38%, the
specificity was 98.1%, and the AUC was 98.74%.At the same period, the ResNet-50 + SVM network
achieved 97.2% accuracy, 97.13% precision, 96.89% sensitivity, 98.67% specificity, and 98.51% AUC.
Furthermore, the AlexNet + SVM network achieved 95.7% accuracy, 96.23% precision, 95.5%
sensitivity, 96.18% specificity, and 99.14% AUC.

Measure GoogLeNet + SVM ResNet-50 + SVM AlexNet + SVM

Accuracy (%) 97.4 97.2 95.7

Precision (%) 97.43 97.13 96.23

Sensitivity (%) 97.38 96.89 95.5

Sepecificy (%) 98.1 98.67 96.18


Results of the Hybrid Feature Deep Learning with the GLCM and LBP Algorithms

Fig 12 highlights the performance of the feature hybrids for analysing CT images for quick diagnosis of cerebral
hemorrhages, namely Google Net deep features + (GLCM and LBP), ResNet-50 deep features + (GLCM and LBP),
and Alex Net deep features + (GLCM and LBP). It attained an accuracy of 98.9%, precision of 99.17%, sensitivity of
98.95, specificity of 99.36%, and AUC of 99.58% when supplied to the ANN utilizing Google Net feature + (GLCM
and LBP). Simultaneously, when supplied to the ANN utilizing ResNet-50 feature + (GLCM and LBP), it achieved
99.1% accuracy, 99.51% precision, 99.32% sensitivity, 99.41% specificity, and 99.62% AUC. Furthermore, when
supplied to the ANN utilizing the Alex Net feature + (GLCM and LBP), it attained an accuracy of 99.3%, precision of
99.36%, and recall of 99.36%. Sensitivity of 99.5%, specificity of 99.57% and AUC of 99.84%.

Figure 10Performance of ANN through hybrid features for early detection of hemorrhages.
Figure 11 shows the performance of the ANN network based on the hybrid features

5.1 Discussion and Comparison of the Proposed Methods

ICH and IVH were split at the same time by CT scans using minimally invasive surgery, alteplase and
clot lyses [23]. A probabilistic technique based on deep Gaussian processes was developed for training by
multi-instance learning and predicting cerebral intracranial hemorrhages. The technique takes the
relationships between features using multiple Gaussian layers that outperform single-layer Gaussian
operations [24]. Numerous patients with neurological problems received head CT scans. The link between
hematoma and clinical factors was demonstrated using a number of univariate and multivariate techniques
[25]. The ResNet-18 model was used to discriminate between normal CT and ICH images. In addition to
using a Grad-weighted class activation mapping method to detect ResNet-18 decisions [26]. Based on
radiology records, a pretrained deep learning network may identify brain haemorrhages. It has been
observed that the effectiveness of fine-tuned deep learning models improves the classification results [27].
Using a three-dimensional deep learning network, the subtypes of hemorrhages are segmented in [48].
The CT images were segmented using the area growth method and then optimized. Three analyzed
strategies for assessing CT images to reach a diagnosis for quick identification of a cerebral hemorrhage
were explored in this study. Each suggestion the system includes many models with varied approaches
and materials. The study's goal was to establish an automated approach that would assist doctors and
radiologists in making a quick diagnosis of a cerebral hemorrhage and determining its location so that the
patient could get appropriate treatment. For all of the proposed procedures, all CT images were enhanced,
all artifacts were eliminated, and the margins of the hemorrhagic region appeared with the same filters.
The first suggested approach review in this paper Google Net, ResNet-50, and Alex Net models, in
which the parameters and training choices were modified to extract and categories deep feature maps. The
overall accuracies of the Google Net, ResNet-50, and Alex Net models were 94%, 91.7%, and 91.5%,
respectively. The second suggested approach is a hybrid strategy composed of two blocks: the first is
CNN models for extracting feature maps, and the second is The SVM technique is used to categories the
feature maps in d block. The overall accuracies of the Google Net + SVM, ResNet-50 + SVM, and Alex
Net + SVM networks were 97.4%, 97.2%, and 95.7%, respectively. The third proposed method which
analyzed in this paper to use an ANN to diagnose the hemorrhage dataset using the hybrid features
extracted by the Google Net, ResNet-50, and Alex Net models, and then uses the PCA algorithm to
reduce the dimensions of the feature maps and combine them with the hybrid features extracted by the
GLCM and LBP algorithms. The ANN network built by merging Google Net feature maps with GLCM and LBP
algorithm features obtained 98.9% total accuracy. The total accuracy of the ANN network based on the coupling of
ResNet-50 feature maps with the features of the GLCM and LBP algorithms was 99.1%. In comparison, the ANN
network achieved 99.3% total accuracy by combining Alex Net feature maps with the features of the GLCM and
LBP algorithms.
System 3
System 1 System 2
Hybrid Model
Deep Machine
Learning GoogLeNet + SVM Learning

GoogLeNet ResNet-50 + SVM SVM()


AlexNet + SVM

ResNet-50 GoogLeNet + (GLCM and RF


LBP)

ResNet-50 + (GLCM and


AlexNet. LBP) ANN AlexNet + MLP
(GLCM and LBP)

MODELS WITH ACCURACY


Conclusion

This systematic review study looked at a variety of CT image proposed models for the rapid detection of
intracranial hemorrhages (ICH), covering a wide range of processes and materials. All of the proposed
methodologies yielded promising results, supporting physicians and radiologists in making decisions
when diagnosing and analyzing hemorrhage images, as well as distinguishing them from suspected
images. This study carefully examines the proposed ML, DL, and hybrid models. The first technique
proposed in this review paper is to use pre-trained CNN models to diagnose the bleeding dataset, namely
Google Net, ResNet-50, and Alex Net, which generated accuracies of 94%, 91.7%, and 91.5%,
respectively. The second approach offered is a Machine Learning model with accuracies of 97.35%,
92.7%, and 97.2%. The third proposed method is a hybrid technique between deep learning (Google Net,
ResNet-50 and Alex Net) to extract the feature maps and a machine learning algorithm (SVM) to classify
the feature maps extracted using deep learning models. Google Net + SVM, ResNet-50 + SVM and Alex
Net + SVM achieved superior results, with accuracies of 97.4%, 97.2% and 95.7%, respectively. The
fourth proposed method is to diagnose the hemorrhage dataset using an ANN algorithm based on the
hybrid features extracted from the Google Net, ResNet-50 and Alex Net models, reduce the feature
dimensions with a PCA algorithm and then combine the features after reducing the dimensions with the
features of the GLCM and LBP algorithms. This method obtained the best results compared with the
other proposed methods, where the ANN reached an accuracy of 99.3%, precision of 99.36%, sensitivity
of 99.5%, specificity of 99.57% and AUC of 99.84% based on the features of Alex Net, GLCM and Lapin
future works, the proposed systems will be evaluated on a dataset containing cerebral hemorrhage types
such as intraventricular, subarachnoid and parenchymal hemorrhages.
References

[1] A. Sage and P. Badura, “Intracranial hemorrhage detection in head CT using double-branch convolutional
neural network, support vector machine, and random forest,” Appl. Sci., vol. 10, no. 21, pp. 1–13, 2020, doi:
10.3390/app10217577.
[2] A. M. Dawud, K. Yurtkan, and H. Oztoprak, “Application of Deep Learning in Neuroradiology : Brain
Haemorrhage Classification Using Transfer Learning,” vol. 2019, 2019.
[3] M. P. Unit, T. Hospital, U. Putra, and H. Sciences, “Asian Journal of Medical Technology (AJMedTech)
Vol. 2 No. 1 (2022),” vol. 2, no. 1, pp. 1–18, 2022.
[4] J.SOFIA, “Analysis of Intracranial Hemorrhage in Ct Brain Images Using Machine Learning and Deep
Learning Algorithm”, [Online]. Available:
https://www.researchgate.net/publication/358460945_Analysis_of_Intracranial_Hemorrhage_in_Ct_Brain_I
mages_Using_Machine_Learning_and_Deep_Learning_Algorithm
[5] X. Gou and X. He, “Deep Learning-Based Detection and Diagnosis of Subarachnoid Hemorrhage,” J.
Healthc. Eng., vol. 2021, 2021, doi: 10.1155/2021/9639419.
[6] A. Tharek, A. S. Muda, A. Baseri Hudi, and A. Baseri Hudin, “Intracranial Hemorrhage Detection in Ct
Scan Using Deep Learning,” Asian J. Med. Technol., vol. 2, no. 1, pp. 1–18, 2022, doi:
10.32896/ajmedtech.v2n1.1-18.
[7] A. Arab et al., “A fast and fully-automated deep-learning approach for accurate hemorrhage segmentation
and volume quantification in non-contrast whole-head CT,” Sci. Rep., vol. 10, no. 1, pp. 1–12, 2020, doi:
10.1038/s41598-020-76459-7.
[8] “Synergic deep learning model–based automated detection and classification of brain intracranial
hemorrhage images in wearable networks,” 2022, [Online]. Available:
https://www.researchgate.net/publication/346536084_Synergic_deep_learning_model-
based_automated_detection_and_classification_of_brain_intracranial_hemorrhage_images_in_wearable_net
works

[9] Awwal Muhammad Dawud, Kamil Yurtkan, Huseyin Oztoprak, "Application of Deep
Learning in Neuroradiology: Brain Haemorrhage Classification Using Transfer
Learning", Computational Intelligence and Neuroscience, vol. 2019, Article ID
4629859, 12 pages, 2019. https://doi.org/10.1155/2019/4629859
[10] He, Kaiming, ZHANG, Xiangyu, REN, Shaoqing, et al. (2016) "Deep residual learning for
image recognition." Proceedings of the IEEE conference on computer vision and pattern
recognition.
[11] Synergicdeeplearning model–based automated detection and classification of
brain intracranial hemorrhage images in wearable networks
CSS Anupama, M Sivaram, EL Lydia, D Gupta… - Personal and Ubiquitous …, 2020 - Springer
[11] ICMLSC '17: Proceedings of the 2017 International Conference on Machine Learning and Soft
ComputingJanuary 2017 Pages 34–39https://doi.org/10.1145/3036290.3Article Source: Brain haemorrhage
detection using a SVM classifier with electrical impedance tomography measurement frames
McDermott B, O’Halloran M, Porter E, Santorelli A (2018) Brain haemorrhage detection using a SVM
classifier with electrical impedance tomography measurement frames. PLOS ONE 13(7):
e0200469. https://doi.org/10.1371/journal.pone.0200469036326
[12] B. C. Ko, S. H. Kim, and J.-Y. Nam, “X-ray image classification using random forests with local wavelet-based
cs-local binary patterns,” Journal of digital imaging, vol. 24, no. 6, pp. 1141–1151, 2011
[13] Wikipedia, Perceptron, https://en.wikipedia.org/wiki/Perceptron, Aug. 2021

[14] Brain Hemorrhage Detection Using Hybrid Machine Learning AlgorithmKhondoker Nazia Iqbal 18101006
Istinub Azad 18101045 Md. Imdadul Haque Emon 18101049 Nibraj Safwan Amlan 18101596 Amena Akter
Aporna 18301236
[15] A smart machine learning model for the detection of brain hemorrhage diagnosis based
internet of things in smart cities
H Chen, S Khan, B Kou, S Nazir, W Liu, A Hussain - Complexity, 2020 - hindawi.com

[16] Senan, E.M.; Jadhav, M.E. Diagnosis of Dermoscopy Images for the Detection of Skin Lesions Using SVM and
KNN. In Proceedings of the the Third International Conference on Sustainable Computing, Jaipur, India, 19–20
March 2021; Springer: Berlin/Heidelberg, Germany, 2022; pp. 125–134

[17]Senan, E.M.; Jadhav, M.E. Techniques for the Detection of Skin Lesions in PH 2 Dermoscopy Images Using
Local Binary Pattern (LBP). In Proceedings of the International Conference on Recent Trends in Image Processing
and Pattern Recognition, Solapur, India, 21–22 December 2018; Springer: Berlin/Heidelberg, Germany, 2020; pp.
14–25.

[18]. Senan, E.M.; Abunadi, I.; Jadhav, M.E.; Fati, S.M. Score and Correlation Coefficient-Based Feature Selection
for Predicting Heart Failure Diagnosis by Using Machine Learning Algorithms. Comput. Math. Methods Med. 2021,
2021, 8500314. [CrossRef] [PubMed]

[19] A. Sage and P. Badura, “Intracranial hemorrhage detection in head CT using double-branch convolutional
neural network, support vector machine, and random forest,” Appl. Sci., vol. 10, no. 21, pp. 1–13, 2020, doi:
10.3390/app10217577.

[20] B. A. Mohammed et al., “Multi-Method Diagnosis of CT Images for Rapid Detection of Intracranial
Hemorrhages Based on Deep and Hybrid Learning,” Electron., vol. 11, no. 15, pp. 1–27, 2022, doi:
10.3390/electronics11152460.

[21] Abunadi, I.; Senan, E.M. Multi-Method Diagnosis of Blood Microscopic Sample for Early Detection of Acute
Lymphoblastic Leukemia Based on Deep Learning and Hybrid Techniques. Sensors 2022, 22, 1629. [CrossRef]

[22]. Al-Mekhlafi, Z.G.; Senan, E.M.; Rassem, T.H.; Mohammed, B.A.; Makbol, N.M.; Alanazi, A.A.; Almurayziq, T.S.;
Ghaleb, F.A. Deep Learning and Machine Learning for Early Detection of Stroke and Haemorrhage. Comput. Mater.
Contin. 2022, 72, 775–796. [CrossRef]

[23]Sharrock, M.F.; Mould, W.A.; Hildreth, M.; Ryu, E.P.; Walborn, N.; Awad, I.A.; Hanley, D.F.; Muschelli, J.
Bayesian deep learning outperforms clinical trial estimators of intracerebral and intraventricular hemorrhage
volume. J. Neuroimaging 2022, 1–9. [CrossRef]

[24]López-Pérez, M.; Schmidt, A.; Wu, Y.; Molina, R.; Katsaggelos, A.K. Deep Gaussian processes for
multiple instance learning: Application to CT intracranial hemorrhage detection. Comput. Methods
Programs Biomed. 2022, 219, 106783. [CrossRef] [PubMed]
[25] Tang, Z.; Zhu, Y.; Lu, X.; Wu, D.; Fan, X.; Shen, J.; Xiao, L. Deep Learning-Based Prediction of
Hematoma Expansion Using a Single Brain Computed Tomographic Slice in Patients with Spontaneous
Intracerebral Hemorrhages. World Neurosurg. 2022, in press. [CrossRef]
[26] Altuve, M.; Pérez, A. Intracerebral hemorrhage detection on computed tomography images using a
residual neural network. Phys. Med. 2022, 99, 113–119. [CrossRef] [PubMed]

[27] Bayrak, G.; Toprak, M. ¸S.; Ganiz, M.C.; Kodaz, H.; Koç, U. Deep Learning-Based Brain Hemorrhage
Detection in CT Reports. Stud. Health Technol. Inform. 2022, 294, 866–867. 48. Inkeaw, P.;
Angkurawaranon, S.; Khumrin, P.; Inmutto, N.; Traisathit, P.; Chaijaruwanich, J.; Angkurawaranon, C.;
Chitapanarux, I. Automatic hemorrhage segmentation on head CT scan for traumatic brain injury using
3D deep learning model. Comput. Biol. Med. 2022, 146, 105530. [CrossRef]

You might also like