IRJET - Detection of Heamorrhage in Brain using Deep Learning
- 1. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 07 Issue: 03 | Mar 2020 www.irjet.net p-ISSN: 2395-0072
© 2020, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 729
DETECTION OF HEAMORRHAGE IN BRAIN USING DEEP LEARNING
AKASH K.1, GAYATHRI M.R2, KARTHIGA.M 3
1,2FINAL YEAR, DEPARTMENT OF BIOMEDICAL ENGINEERING, ANNA UNIVERSITY, CHENNAI, INDIA
3ASSISSTANT PROFESSOR, DEPARTMENT OF BIOMEDICAL ENGINEERING, ANNA UNIVERSITY, CHENNAI, INDIA
-----------------------------------------------------------------------***--------------------------------------------------------------------
Abstract- Cerebrovascular diseases are the third cause of
death in the world after cancer and heart diseases. Brain
heamorrhage is one of the most common cerebral vascular
diseases. Brain heamorrhage is caused by the bursting of
brain artery leading to bleeding and can have a fatal
impact on brain function and its performance. For
diagnosis of heamorrhage medical experts suggest either
MRI or CT .CT images are used in greater ratio due to its
ease of use, price constraints and high speed. The
identification of cerebral heamorrhage is not known
immediately. Therefore we need a certain method that can
segment the CT scan image quickly and automated. The
goal is to obtain the segmentation of brain part that is
affected with heamorrhage quickly and accurately using
the method of Deep Learning. So patients with cerebral
heamorrhage can immediately obtain the medical
treatment in accordance with the needs.
Keywords: Brain heamorrhage, CT images, deep
learning
1. INTRODUCTION
The brain is one of the largest and most complex organs
in the human body. It is made up of more than 100
billion nerves that communicate in trillions of
connections called synapses. The brain integrates
sensory information and direct motor responses. It also
helps the people to think, feel, and emote. Thus brain is
called the Centre of Learning which gives commands to
all other organs in the body. There are many situations
where the brain gets affected, infected, injured so that
their normal activity gets collapsed. One of those
situations is the development of heamorrhage. Brain
heamorrhage is a serious category of head injury that
can have a fatal impact on brain function and
performance. Brain heamorrhage can be diagnosed by
two kinds of imaging modality: Computed Tomography
(CT) and Magnetic Resonance Imaging (MRI). After
going through many of the literatures and checking with
medical experts CT images are chosen in this work. CT
images are known to have many advantages over MRI
such as: wider availability, lower cost and higher speed.
Moreover, CT scanner might be favored over MRI
scanner due to patient-related issues such as the patient
being too large to fit in the MRI scanner, claustrophobic,
has metallic or electrical implants or is unable to remain
motionless for the duration of examination due to age,
pain or medical conditions . Finally, the quality of CT
images is high enough to accurately diagnose brain
heamorrhage.
Fig 1 Survey reports of people affected with
heamorrhage globally
2. RELATED WORK
There were many approaches related to detection of
heamorrhage. [1] Alexandra Lauric and Sarah Frisken
proposed soft segmentation methods like Bayesian
classification, Fuzzy c-Means, and Expectation
Maximization is applied on CT brain images and they
have compared all these methods to produce a best
accuracy. The first method used a Bayes rule to predict
that a given pixel belongs to a particular class by using a
conditional probability. The second segmentation
alternates between partitioning the pixels into clusters
and updating the parameters of each cluster. Like many
clustering algorithms, FCM favors large clusters and, as a
result, pixels belonging to small clusters are often
misclassified. To compensate the misclassification, they
used the Population-Diameter Independent (PDI)
algorithm, which was introduced in [2] as a variation of
FCM. The PDI algorithm uses cluster weights to balance
the influence of large clusters. The third segmentation
method partitions pixels into clusters by determining the
maximum likelihood parameters of a mixture of known
distributions. All three methods perform segmentation
by constructing statistical models but they have both
strengths and limitations. Bayesian classification is
simple, fast and robust, but it requires training and is
sensitive to the accuracy of training data. FCM is an
efficient, self-organizing clustering algorithm which does
not require training. However, unlike the Bayesian
classifier, FCM does not explicitly incorporate prior
knowledge. Expectation Maximization combines the
strengths of both algorithms, it is based on Bayes rule, so
it incorporates prior knowledge, but it is an
- 2. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 07 Issue: 03 | Mar 2020 www.irjet.net p-ISSN: 2395-0072
© 2020, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 730
unsupervised segmentation method, so it does not
require any training. [3]The authors used a
aforementioned approach like image preprocessing,
image segmentation, feature extraction, and
classification to detect whether a brain heamorrhage
exists or not in a Computed Tomography (CT) scans of
the brain which is a binary classification. Moreover, the
type of the heamorrhage is identified which is a multi
classification .The accuracy is 92% for the binary
classification and this can be improved with large
number of datasets and better feature extraction
algorithm .[4] In this article the authors used a
watershed segmentation and Multi Layer Perceptron to
find the presence of heamorrhage. Features are
extracted using watershed segmentation and Gabor
filter. The extracted features are classified using
Multilayer Perceptron (MLP). Finally images are
classified as stroke and non-stroke images.[5,8] The
author’s goal is to segment the part of brain bleeding
more quickly and accurately. The preprocessing of CT
scan image starts from color filtering, erosion and
dilation methods to eliminate the noise contained in the
image. Then they performed the watershed and cropping
segmentation to separate the skull bones of the CT
image. Median filter is used to improve the image
quality. Then the image is again segmented using the
threshold method to separate the image of cerebral
heamorrhage as the observed object. At last the
calculation of area and volume percentage of bleeding in
the brain is performed. From this approach the
calculation of brain area has an average error of 1.13%
and the calculation of the bleeding area has an average
error of 11.17%. This system can be improved with the
incorporation of better feature extraction and pre-
processing methods to improve the accuracy rate.[6,7]
Detect and classifies stroke in skull CT images through
analysis of brain tissue densities. Featuring techniques
such as gray level co-occurrence matrix, local binary
patterns, central, statistical, Hu’s, Zernike’s moments
were used.[11]The author focused on detecting the
correct location and type of the heamorrhage in MRI
Brain image. To segment the hemorrhagic region,
structure specific Multi level Set evolution algorithm is
implemented. To extract sharpened tetra features an
enhanced Local tetra pattern based feature extraction
algorithm and the features are optimized by applying an
enhanced Grey Wolf Optimization algorithm. Finally, a
Relevance Vector Machine based Classifier is
implemented to classify the types of the heamorrhages.
3. WHY DEEP LEARNING?
Automated image analysis is a tool based on machine
learning algorithms .Machine learning are the key
enablers to improve the quality of image diagnosis and
helps in interpretation by facilitating through efficient
identification. Machine learning includes simple Neural
Network, Deep Learning, Artificial neural networks
which are structurally and conceptually inspired by
human biological nervous system. Deep learning is the
one extensively applied technique that provides state of
accuracy. It is the most effective and supervised machine
learning approach that uses deep neural network
model(DNN) which is different from the simple neural
network. DNN shows a larger approximation to human
brain as compared to simple neural network.
4. STRUCTURE OF DEEP NEURAL NETWORK
The basic computational unit in a simple neural network
is the neuron which takes multiple signals as inputs,
combines them linearly using weights, and then passes
the combined signals through nonlinear operations to
generate output signals. Perceptron is one of the earliest
neural network methods. It consists of input layer that is
directly connected to output layer and was good to
classify linearly separable patterns. To solve more
complex pattern, neural network was introduced that
has a layered architecture i.e., input layer, output layer
and one or more hidden layers.
Fig 2 The structure of deep neural network
Deep Neural network consist of interconnected neurons
that sums up the input data and apply the activation
function to the summed data and finally provides the
output that might be propagated to the next layer. Thus
adding more hidden layer allows to deal with complex as
hidden layer capture nonlinear relationship. Deep
learning will not only help to select and extract features
but also construct new ones, furthermore, it does not
only diagnose the disease but also measures predictive
target and provides actionable prediction models to help
physician efficiently.
5. VARIOUS ALGORITHMS OF DEEP LEARNING
Various types of deep learning algorithms which are in
use are Convolutional neural networks (CNN), Deep
belief network (DBN), Deep auto encoder (DA), Deep
Boltzmann machine (DBM), Deep conventional extreme
machine learning (DC-ELM), Recurrent neural network
(RNN). Deep learning based algorithms showed
promising performance as well speed in different
domains like speech recognition, text recognition, lips
reading, computer-aided diagnosis, face recognition,
drug discovery.
- 3. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 07 Issue: 03 | Mar 2020 www.irjet.net p-ISSN: 2395-0072
© 2020, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 731
6. PROPOSED METHOD
Fig 3 Block diagram
Among the above mentioned algorithms CNN model is
widely used in digital imaging processing because it
reduces input image size without any loss of information
and it also helps to improve the computational speed.
CNN consists of several steps they are Input Layer,
Hidden Layer, Activation Function, Max pooling Layer,
Dense layer and Drop out Layer. CNN works under
Sequential process which means Hidden layer output
will be input to Activation function, output of Activation
function will be input to next layer.. The count of images
are 100 normal and 100 abnormal (heamorrhage
affected brain) images. Among them the first 90 normal
and abnormal brain images is used for training and the
remaining 10 normal and abnormal images is used for
testing which are in the ratio of 8:2.
TABLE 1 the training and testing dataset
Type of image Training Testing
Normal brain 90 10
Abnormal brain 90 10
Total 180 20
Priorly images of Haemorrhagic Brain and Normal Brain
are labelled in separate CSV file for dependent variable
where normal brain is labelled as ‘0’ and hemorrhagic
brain is labelled as ‘1’.we add 3 Layers with the dense of
32, and 6 layers with the dense of 64 to compare the
accuracy with a kernel size of 3*3 and strides 2 ,and Relu
as a Activation Function to get maximum possibility of a
positive integer. Convolutional layer will continue for 32,
64,512 to prevent the loss of information and finally a
web development model is created to test the results.
Algorithm 1 for The Convolution model
1. model.add(Conv2D(32,kernel_size=3,strides=2,padding='same'activation='relu',
input_shape=input_shape))
2. model.add(MaxPooling2D(pool_size=2))
3. model.add(Conv2D(32, kernel_size=3, strides=2, padding='same', activation='relu'))
4. model.add(MaxPooling2D(pool_size=2))
5. model.add(Conv2D(64, kernel_size=3, strides=2, padding='same', activation='relu'))
6. model.add(GlobalAveragePooling2D())
7. model.add(Dropout(0.4)) model.add(Dense(64, activation='relu'))
8. model.add(Dropout(0.4))
9. model.add(Dense(1,activation='sigmoid’))
10. return model
- 4. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 07 Issue: 03 | Mar 2020 www.irjet.net p-ISSN: 2395-0072
© 2020, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 732
7. RESULTS
Input image will be uploaded, back-end process takes
the input and predict with trained model. Depending
upon the threshold value Haemorrhage and Normal
Brain will be identified and displayed.
Fig 5 the brain with heamorrhage
Fig 6 the normal brain predicted
Fig 7 Number of Epochs and the accuracy resulted
With this method the heamorrhage in brain can be
predicted with the accuracy of 94.4% in the web
development of python. Additionally, we can use his
method to predict the types of heamorrhage and to find
the area of heamorrhage in brain to grab a better
performance.
REFERRENCES
[1]Alexandra Lauric and Sarah Frisken “Soft
Segmentation of CT Brain Data” in January ,2007
ResearchGate
[2] Shihab AI.: Fuzzy Clustering Algorithms and Their
Application to Medical Image Analysis. In Ph.D. thesis,
University of London 2000
[3]Mahmoud Al-Ayyoub, Duaa Alawad, Khaldub Al-
Darabsah & Inad Alkarahh ‘Automatic Detection and
Classification of Brain Heamorrhages’ published on 11
July 2015.
[4] Amutha Devi C, Dr, S. P. Rajagopalan proposed their
idea in “Brain Stroke Classification Based On Multilayer
Perceptron Using Watershed Segmentation and Gabor
Filter” in Journal of theoretical and Applied Information
Technology,2013.
[5] Rizal Romadhoni Hidayatullah, Riyanto Sigit, Sigit
Wasista proposed their idea in the paper “Segmentation
of Head CT-Scan to Calculate Percentage of Brain
Heamorrhage Volume” published in IEEE 2017.
[6] Chiun-Li Chin, Bing-Jhang Lin, Guei-Ru Wu, Tzu-Chieh
Weng, Cheng-Shiun Yang, Rui-Cih Su, Yu-Jen Pan “ An
Automated Early Ischemic Stroke Detection System
using CNN Deep Learning Algorithm” in Icast201, 2017
[7] Chiun-Li Chin, Bing-Jhang Lin, Guei-Ru Wu, Tzu-Chieh
Weng, Cheng-Shiun Yang, Rui-Cih Su, Yu-Jen Pan “ An
Automated Early Ischemic Stroke Detection System
using CNN Deep Learning Algorithm” in Icast,2017
[8]C.S.Ee, K.S.Sim, V.Teh, F.F.Ting “Estimation of Window
Width Setting for CT Scan Brain Images Using Mean of
Greyscale Level to Standard Deviation Ratio”
[9]Elizˆangela de S. Rebouc¸as, Alan M. Braga, R´oger
Moura Sarmento,Regis C. P. Marques and Pedro P.
Rebouc¸as Filho “Level set based on brain radiological
densities for stroke segmentation in CT images” in 30th
International symposium on Computer based Medical
Systems,2017
[10]Igor Bisio, Alessandro Fedeli, Fabio Lavagetto,
Matteo Pastorino, Andrea Randazzo, Andrea Sciarrone,
Emanuele Tavanti “Mobile Smart Helmet for Brain
Stroke Early Detection through Neural Network based
Signals Analysis” in IEEE Global Communications
Conference , 2017
[11] Nita Kakhandaki, S. B. Kulkarni “A Novel Framework
for Detection and Classification of Brain Heamorrhage”
IJRTE Journal Volume-7, November 2018