Determination with Deep Learning and One Layer Neural Network for Image Processing in MultiSlice CT Angiogram

Sarvadeva BhatlaMurali Krishna. Int. Journal of Engineering Research and Application www.ijera.com
ISSN : 2248-9622, Vol. 7, Issue 4, ( Part -6) April 2017, pp.01-07
www.ijera.com DOI: 10.9790/9622-0704060107 1 | P a g e
Determination with Deep Learning and One Layer Neural
Network for Image Processing in MultiSlice CT Angiogram
Sarvadeva BhatlaMurali Krishna*, Dr. M Chandra Shekar**
*(PhD Scholar in ECE, at Rayalaseema University with Regd No. PP-ECE-0044, AP, India)
** (Senior Manager Bharath Dynamics Limited, Hyderabad, Telangana, India)
ABSTRACT
Today’s world Coronary artery disease is the most common cause of death worldwide and thus early diagnosis.
Well-timed opportune of this disease can lead to significant reduction in its morbidityand mortality in both
younger and older for angiogram test. In this research multi slice CT scanner is used for heart angiogram test.
With the help of this multi slice CT angiogram image we detect the hart diseased or not. For this disease
identification and classification of angiogram images many machine learning algorithms are previously
proposed those are SVM RBF and RBF neural network. Problem with SVM isnon-liner method when use any
type of application will miss most liner ways of blood vessels and lack of speed in process. For non linear
classification we are using RBF SVM. Problem with RBF neural network is not solve the hierarchal and
component based problems, so resolve the problem using deep learning. This issue drastically improves the
estimation efficiency for real time application. This methodology consumes less time for both learning as well
as testing comparatively than any other methods. This issue highly improves the estimation efficiency and
accuracy for real time 256, 512 slices CT scan angiogram image.
I. I
NTRODUCTION
Most of the human beings are suffering
Heart diseases. Some cases identification of the
hart dices in particular blood vessel is difficult so,
in these cases doctors suggested the angiogram test
with the help of multi slice CT scan. A CT scanner
is the same way as a conventional x-ray but instead
of taking one image a CT scanner takes multiple
images, or slices. A computer program gathers all
the images and compiles them to create a three
dimensional image of the internal structures being
examined. Advances have improved the sensitivity
and usefulness of CT scanners since they were
developed. It can provide a 3 dimensional image of
an internal structure, it can detect differences in
tissue density, and the x-rays can be focused
directly on specific areas, producing a clear image
[1] [2].In this research we are using advanced CT
scanners that are 256 and 512. Basic CTA scan 3D
image is shown below Figure 1.The basic heart
images of patient with coexistence of transvenous
(active) and epicedial (not active) leads; A,
B. Three-dimensional images of the heart; * -
visible patch electrodes; C1, C2. Diagnostic
visualization of the coronaries: left anterior
descending artery (LDA), right coronary artery
despite (RCA) the presence of both types of
lead; D. Multi planar reformatted reconstruction
with visible artefacts from the leads[3].
Figure1: Basic CTA Scan 3D Image
In process of medical analysis CT images
used normally first step is segmentation. The
research for medical image analysis accurate
angiogram segmentation is a crucial prerequisite
for computer-aided hepatic disease diagnosis and
treatment planning. In this research Multi slice CT
angiogram segmentation is performed with the help
of k-means++ clustering algorithm [4].Figure 2
represents the Coronary artery disease multi slice
CT scanner image with segmentation. A Curved
multi-planar reformations (CMPR) of the right
coronary artery (RCA), Maximal
inspiratorypressure (MIP), and perpendicular cut-
RESEARCH ARTICLE OPEN ACCESS

planes (Inlays a–d) delineate multiple non-calcified
plaques in mid-left anterior descending (LAD), one
leading to a haemodynamically significant stenos is
(asterisk, cut-plane b). In the RCA, CMPR B,
Virtual reality (VR) image, and MIP demonstrate
multiple non-significant wall irregularities in the
mid-segment (arrows) and occlusion of the distal
segment (arrowheads) [5]. As shown in the internal
carotid artery (ICA) C and D, all lesions were
correctly identified.
Figure 2: segmentation of CT angiogram heart
image
This heart angiogram multi slice CT
image Classification and retrieval is the more
difficult job from fast decade, because of the
exponential growth of images. The accuracy
assessment of images correspondence using deep
learning all are emphasized in this research [6]. All
aspects of images features are covered. A total of
11 features are defined in four categories: spatial
pixel level, Texture, shape and relational. In this
research, all aspects of features are extracted and
analyzed on different Data sets.
After extracting the features then
identifies the angiogram heart image is diseased or
not. For that olden days use the machine learning
algorithms that are support vector machine and
RBF neural network. Here SVM is Non-Liner
method when use any type of application will miss
most liner ways of blood vessels and lack of speed
in process. For Non-Linear classification we are
using RBF SVM. A common disadvantage of non-
parametric techniques such as SVMs is the lack of
transparency of results. RBF neural network is not
solving the hierarchal and component based
problems, in order to overcome this problem by
using Deep Learning. This issue extremely
improves the estimation efficiency of real time
application. Final analysis of the research gives
more improve of the accuracy [7] [8].
II. BACK GROUND
Nowadays, cardio- and cerebro-vascular
diseases have greatly threatened human health.
Established invasive test for non invasive coronary
artery evaluation with high sensitivity.Cardiac
automatic disease is identified with the help of
angiogram test by using multislice CT scanner.
These can produce an image in less than a
second and thus can obtain images of the heart and
its blood vessels (coronary vessels). The First
Multislice CT invention by Kalender in the 1980s,
helical scan CT machines has steadily increased the
number of rows of slices they deploy [9] [10]. In
this research we are using advanced CT scanners
that are 256-slice and 512-slice.
The 256 CT scanners have craniocaudal
coverage of approximately 12 and 16 cm,
respectively. This potentially allows the heart to be
scanned in one tube rotation and one heartbeat,
without table movement. 512-slice CT scan has
resolutions of 512x512 pixels With varying pixel
sizes and slice thickness between 1.5 − 2.5 mm,
acquired on Philips and Siemens Multi-Detector
CT scanners with 120 kVp tube voltage [11]. This
heart angiogram Multislice CT Image
Classification and retrieval is the more difficult job
from fast decade, because of the exponential
growth of images [12]. The accuracy assessment of
images correspondence using Deep Learning all are
emphasized in this research.
III. PROPOSED METHODOLOGY
Figure 3: Flow chart of proposed methodology
architecture
Major process divided into four steps:

 The initial step is data preparation;Data
analysisin this first collecting the data from
data sets after that this data is pre-processed
before data selection.
 Subsequently data processing the entire data is
grouped into different datasets. Each data set
consists of Train Data sets and Test Data sets.
Test and Train data set whose details are
explained in the next sections.
 Next followed step is segmentation to extract
Feature from trained and testdata sets.
 The heart Step for this research is Deep
Learning system.
 After selecting the method build the system
first and then test with different test cases for
evaluating the performance in terms of
accuracy.
IV. TEST DATA
In this Research, data werecollected from
Agecanonix, FIVIX and MAGIX are the most
widely used data sets for many research problems.
This datasets consisting of different kinds of
Multislice CT Heart.Angiogram images. The
images are of the size 128X128, 256X256 and
512X512. This research used the Agecanonix,
FIVIX and MAGIX image databases, which is
consisting of 1106, 1754, 760 Images. Each data
set consisting of minimum three subjects.Each
subject consisting of 50 to 100 angiogram images.
Agecanonix data set is contains13
subjects. First subject consisting of 6 records and
355 images. From second subject to ninth subject
contain equal amount of images, but different no.
of records. 10 and 11 subjects contain different no.
of records, but equal no. of images. The entire
Agecanonix data set is represented in below bar
chart figure 4.
Figure 4: Agecanonix Data Set Bar C Hart
FIVIX data set is contains 3 subjects. First
subject contain 11 records and 414 images. Second
subject contains 13 records and 530 images. Final
subject contain more no. Of records 404 and more
images 810. Entire 3 subject’s relation is shown in
below figure 5.
Figure 5: FIVIX Data Set Bar Chart
MAGIX data set is contains 10 subjects.
All the subjects’ records are different, but it
contains equal no. of images. It shown in below bar
chart 6.
Figure 6: MAGIX Data Set Bar Chart
In this research 3 data sets are using .each
data set is containing multiple no of angiogram
heart images. After that extract the feature sets
from these data sets.

Technique Name Feature Name
Statistical Pixel-Level Mean
Statistical Pixel-Level Variance
Statistical Pixel-Level Gray Value
Statistical Pixel-Level Contrast of Pixel
Statistical Pixel-Level Edge Gradient
Shape Circularity
Shape Compactness
Shape Moments
Shape Chain Codes
Relational Relational Structure
Relational Hierarchical Structure
Table 1: Feature Extraction Descriptions
V. RBF NEURAL NETWORKS
The Radial Basis Function Neural
Network (RBF NN) is a three layer design with
weights and bias optimization values.Generally in
Radial Basis Function, input layer consists of input
data which is extracted from image features. This
work used different dimensions of the image
features like statistical pixel level, Texture, shape
and Relational.In the hidden layer, it contains one
additional node than input node. This layer consists
of Centroids which are calculated by using
Gaussian distribution function internally. Outputs
are computed from processing layer and hidden
layer plus Bias values.It is represented in below
equation 1(13).
Output Layer =Sum of (Hidden Layer * Weights) +
Bias value. (1)
Figure 7, shows the block diagram of RBF Neural
network. The input, output and hidden layers are
J1, J2, J3 neurons respectively.
Above equation 2 represents the bias in the output
layer .below equation 3 represents the non linearity
at hidden nodes
Figure 7: Block diagram of RBF neural network
Whereyi(xˉ)is the ith output, wki is the
connection weight from the kth hidden unit to the ith
output unit, and denotes the Euclidean norm.
The RBF is typically selected as theGaussian
function, and such an RBF network is usually
termed the Gaussian RBF network. The output
function f compute by using the below
approximation function like below equation 4 (14).
......
(4)
In this research Neural Network and Deep
Learning are proposed for the learning and
classification of Multi Slice CT Angiogram images.
But RBF neural network is not solving the
hierarchal and component based problems, in order
to overcome this problem by using Deep Learning.
This issue extremely improves the estimation
efficiency of real time application (15).
VI. DEEP LEARNING
In Neural networks good weights learning
algorithm is RBF NN with1 hidden layer,
explained in. Not good at learning the weights for
networks with more hidden layers. The problem
may not solve one layer, but our research
requiredfor Multi types of futures to resolve the
problem. Because blood vessels is not signal types
of features extraction (16).
A Deep Neural Network is an Artificial
Neural Network with multiple hidden layers of
units between the input and output layers. DNN is
used in computer vision. Deep Learning algorithms
for Unsupervised or Semi Supervised feature
learning and hierarchical feature extraction
algorithm.
The first Supervised Deep Learning was
invented by Ivakhnenko and Lapa in 1965.
Artificial Neural Networks was introduced by Igor
Aizenberg and colleagues in 2000. Unsupervised
representation learning methods such as Restricted
Boltzmann Machines (RBM) may outperform
standard filter banks because they learn a feature
description directly from the training data(17).

Figure 8: Deep Learning
Like this above figure 7 each layer our research
different future data sets with layer output
VII. ACCURACY ASSESSMENTS
In this paper of research are used three
different data sets. About data sets are specified in
section 4.Each data set, divided into some set of
data has training data set and left our part is test
data set. Each data set contained the heart
angiogram images this data is used to identify the
diseased blood vessels in the human heart. Instance
of data set is already defined how much blood
vessels part of data sets for train data. Has first
training data set is defined 90% data set from
original data and 10% consider has test data set.
Like that, decrease training data set % and increase
test data set % with different 7 cases are
Table 2: Training and Test Data Sets % with
Different Test Cases
Used with above training and test data sets
for each analysis division of % shown in above
Table 2. Assessment of accuracy done for
following methods with above train and test data
set cases.Those are shown below table 3
 RBF neural network
 Deep Learning.
Cases Accuracy of RBF NN
Case #1 96%
Case #2 95%
Case #3 95%
Case #4 95%
Case #5 90%
Case #6 86%
Case #7 85%
Table 3: Accuracy of RBF NN
Figure 9: Bar Chart for Accuracy of RBF NN
Cases Accuracy of Deep
Learning
Case #1 98%
Case #2 97%
Case #3 96%
Case #4 96%
Case #5 93%
Case #6 90%
Case #7 86%
Table 4: Accuracy of Deep Learning
Seq. No
Case
Training Data
Set % from
Original Data
Set
Test Data
Set % from
Original
Data Set
Case #1 90 10
Case #2 80 20
Case #3 70 30
Case #4 60 40
Case #5 50 50
Case #6 40 60
Case #7 30 70

Figure 10: Accuracy of Deep Learning
Figure 11: Bar Chart for Acc Comparison between
RBF NN and Deep Learning
Accuracy is exhibited above Bar Chart for
Acc Comparison BW RBF NN and Deep Learning
for research methodologies. Clear proved Deep
Learning is more accurate compare than RBF NN
and another thing is here clear specified even
decrease train data set also accuracy is stabilized in
all aspect not much decrease as like RBF NN [18].
Additional analyzed RBF NN stabilized
different with Weights and Bias values. Total
weight values are 132 values and 2 bias values.
VIII. CONCLUSIONS
This paper, proposed a method to find the
Coronary Heart diseases. In this research using
Computed Tomography (CT) scanner for identified
blood vessel in heart area. The problem of
identification blood vessels in heart is binary
classification. Majorly binary classification
algorithmis SVM. SVM is the Non- linear
classification, so it misses more no. of linear blood
vessels. In order to overcome this problem
implements the RBF NN. It is the both linear as
well as Non-linear classification.RBF neural
network is not solving the hierarchal and
component based problems, in order to overcome
this problem by using Deep Learning. Hence, this
paper gives clear information about improved
classification using deep learning compare with
existing methodology of RBF neural network.
Future extended with advanced
methodology for image classification, feature
extractions and best accuracy assessment.
REFERENCES
[1]. Lei Z, Han P, Xu H, Yu J. Application study
of 512-slice computed tomography low dose
coronary angiography. Digit Med 2015;
1:28-33.
[2]. Sun ZH, Cao Y, Li HF. Multislice computed
tomography angiography in the
[3]. diagnosis of coronary artery disease.J
Geriatr Cardiol. 2011 Jun;8(2):104-13.
doi:10.3724/SP.J.1263.2011.00104.
[4]. Thayssen P, Jensen LO, Lassen JF, et al. The
risk and prognostic impact of definite stent
thrombosis or in-stent restenosis after
coronary stentimplantation.
EuroIntervention 2012;8(5):591–8.
[5]. Rixe J, Achenbach S, Ropers D, et al.
Assessment of coronary artery stent
restenosis by 64-slice multi-detector
computed tomography. Eur Heart J
2006;27(21):2567–72.
[6]. Jensen JK, Jensen LO, Terkelsen CJ, et al.
Incidence of definite stent thrombosis or in-
stent restenosis after drug-eluting stent
implantation for treatment of coronary in-
stent restenosis. Catheter Cardiovasc Interv
2013;81(2):260–5.
[7]. Pugliese F, Cademartiri F, van Mieghem C,
et al. Multidetector CT forvisualization of
coronary stents. Radiographics 2006;26
(3):887–904.
[8]. Jung JI. Multidector CT imaging of coronary
artery stents: is this method ready for use?
Korean Circ J 2007;37(11):521–9.
[9]. Quoc, L., Ngiam, J., Chen, Z., Chia, D.,
Koh, P. W., and Ng, A. Tiled convolutional
neural networks. In NIPS 23. 2010.
[10]. Vinyals, O., Jia, Y., Deng, L., and Darrell,
T. Learning with Recursive Perceptual
Representations. In NIPS, 2012.
[11]. Xu L, Sun Z. Virtual intravascular
endoscopy visualization of calcified
coronary plaques: a novel approach of
identifying plaque features for more accurate
assessment of coronary lumen stenosis.

Medicine. 2015; 94(17):e805. doi:
10.1097/MD.0000000000000805 PMID:
25929936; PubMed Central PMCID:
PMC4603061.
[12]. Almoudi M, Sun Z. Coronary artery calcium
score: Re-evaluation of its predictive value
for coronary artery disease. World journal
ofcardiology. 2012; 4(10):284–7. doi:
10.4330/wjc.v4.i10.284 PMID: 23110244;
PubMed Central PMCID: PMC3482621.
[13]. Syeda-Mahmood, T., Wang, F., Beymer, D.,
Amir, A., Richmond, M., Hashmi, S. Aalim.
“Multimodal mining for cardiac decision
support”. Comput. Cardiol. 34, 209–212
(2007).
[14]. K.-L. Du and M. N. S. Swamy, Neural
Networks in a Softcomputing Framework,
Springer, London, UK,2006.
[15]. Hofman MBM, Paschal CB, Li D, et al. MRI
of coronary arteries: 2D breath-hold vs 3D
respiratory-gated acquisition. J Comput
Assist Tomogr.
[16]. Saman Sarraf, and Ghassem Tofighi,
"Classification of Alzheimer's Disease
using fMRI Data and Deep Learning
Convolution Neural Networks," arXiv
preprint arXiv: 1603.08631,2016.
[17]. S. S. Watkins and P. M. Chau, “A radial
basis function neuron computer
implemented with analog VLSI circuits,” in
Proceedings of the International Joint
Conference on Neural Networks, vol. 2, pp.
607– 612, Baltimore, Md, USA.

Determination with Deep Learning and One Layer Neural Network for Image Processing in MultiSlice CT Angiogram

More Related Content

What's hot (20)

Similar to Determination with Deep Learning and One Layer Neural Network for Image Processing in MultiSlice CT Angiogram (20)

Recently uploaded (20)

Determination with Deep Learning and One Layer Neural Network for Image Processing in MultiSlice CT Angiogram