deep learn about blood vessel auto1.pptx

GOPALAN COLLEGE OF ENGINEERING AND
MANAGEMENT
BANGALORE-560 048
Department of Computer Science and Engineering
Project Review
On
Automated Alignment of Blood
Vessel in Retinal Fundus Images
Submitted By
Bharathi.R (1GD16CS006)
Priyadarshini.B (1GD16CS040)
Sowmya.M (1GD16CS053)
Suman Kumari Yadav (1GD16CS054)
Under the guidance of
Dr.V.N.Manju, B.E, M.Tech, PhD
Assistant Professor of CSE ,
GCEM Bangalore-048.

CONTENT OUTLINE
 Title
 Abstract
 Introduction
 Objective
 Existing System
 Proposed System
 Literature Survey
 Software andHardware Requirements
 Design
 Implementation
 Snapshots
 Conclusion andFuture Enhancement
 References

ABSTRACT
 Optimal transport supports modern imageprocessing applications
such as
• medical imaging
• scientific visualization.
 Enables greatflexibility in modeling problems relatedto image
registration, asthe choice of suitablematchingmodels to alignthe
images.
 Automated framework for fundus imageregistration which unifies
optimal transport theory,
 Imageprocessing tools andgraph matchingschemes into a
functionalandconcise methodology.
 Given two ocular fundus images,we construct representative
graphs which embed in their structures spatialandtopological
information from the eye’sblood vessels.

 Medical imaging has an important component in disease
diagnosis.
 The imaging also helps in planning ,carrying out and evaluating
surgical and radiotherapeutic procedures.
 Medical image modalities can be classified into two types
Functional and anatomical
 Functional imaging is done to understand the functions of the
underlying morphology. Examples of functional image are
 SPECT(Single-Photon Emission Computed Tomography)
 PET(Positron emission Tomography)
 fMRI(Functional MRI)
 Anatomical Imaging is mainly to capture morphology of an
organs. Examples of anatomical imaging are x-ray, CT scan, MRI
scan, Ultra sound etc.

INTRODUCTION
 In 2012, the World Health Organization(WHO)
estimated that
 39 million people in theworld are blind
 285 million arevisually impaired
 246 million havelow vision degree.
 Glaucoma,the second leadingcause of blindness
worldwide.
 Damages theoptic nerve. It occurs when afluid(calledaqueous) builds up in
the front part of the eye,increasing thepressure on it.

 Image registration is a solution to is a solution to this problem
where information from two or more images is fused to get a
single image with more details for disease diagnosis.
 In this work we survey the current solutions for medical image
registration and identify the advantages and disadvantages of
retinal fundus images.
 Retinal image scans are used to identify a most serious disease
affecting eye called as Glaucoma.
 Glaucoma is of two types –
 Open angle
 Close angle

Open angle image Close angle image

• As the medical diagnosis for glaucoma and other
eye disorders,
• The use of image processing algorithms became
necessary, especially when ophthalmologists need
to manage a large set of fundus images.
• In our project
• We present an automated registration framework for
aligning blood
• vessels in retinal fundus images.
• the present approach relies on the stable theory of OT
• OT in conjunction with graph-based models to
precisely match retinal blood vessels. The
performance and usefulness of the system is
evaluated.

• Vessel-based techniques attempt to generate a tree-like model
representing the vascular structures of the retina so that the
topological maps of these structures are detected and
captured.
• Vessel Optimal Transport for fundUS image alignment, retains
some desirable characteristics when registering images such as
• registration accuracy for pairs of images contaminated by specular
noise
• high quality performance for high-resolution retinal images
• stability with respect to the systematic registration task.
• VOTUS, output can enhance the perception of anatomy
changes between the retinal images so that it may assist
ophthalmologists in preventing and diagnosing eye diseases
and other related impairments.

 The objectives of this project are
 To design aregistration scheme to fuse
two ocular images.
 Automated imageregistration framewrok.
 Demonstrate the accuracy and
effectivenessof the proposed registration
scheme.
OBJECTIVE

EXISTING SYSTEM
The existing solutions are grouped into three categories
◦ Vessel based
◦ Intensity based
◦ Key point based
Vessel-based, relies on graph-based representations of the
eye’s retina to accomplish the registration
Intensity-based one performs the eye’s allignment by
exploiting specific attributes of the images such as colors,
contrasts, and gradients.
Key point based aims at establishing a correspondence
between sets of keypoints extracted from the retinal images.

PROPOSED SYSTEM
• Our methodology provides a new facet of OT
o As a constrained graph matching problem to align vascular
vessels on retinal images.
o Graph representatives are constructed from the pair of
the acquired images
• The present approach relies on stable theory of OT in
conjunction with graph-based models to precisely match
retinal blood vessels.
• So that their nodes are viewed as a multi-valued set
of features and evaluated as key points to achieve
the registration.

 Once the key points are obtained,
 We compute their direct correspondence between the images, by
solving the proposed OT problem.
 The designed optimization model allows us to establish the
matching from a customized cost function that penalizes outliers as
long as the matches are determined.
First, the flexibility of VOTUS to cope with different kind of
images is shown in the below figure.
 In contrast, VOTUS achieves more consistent and pleasant
results in all cases, mainly regarding the quality of matching
refinement as shown by a large amount of white color in the
resulting montages.
 We can observe from the results that VOTUS performs best
when compared with other evaluated methods.

Visual results generated by VOTUS

Paper Name Author Year Limitations
VesselOptimalTransport
forAutomatedAlignment
of Retinal Fundus Images
Danilo Motta
Wallace Casaca
Afonso Paiva
2019
Sensitiveto color
variations
Gaussian fieldestimator
with manifold
regularizationfor retinal
image registration
J.Wang,J.Chen,H.
Xu, S.Zhang
2019
The computational
complexity is high.
A vascularimage
registration method
based on network
structure andcircuit
simulation
L.Chen,Y.Lian,Y.
Guo,Y.Wang,T.S.
Hatsukami
2017
The vessel
information must be
present in allthe
images
LITERATURE SURVEY

Paper Name Author Year Limitations
Retinal image
registration under the
assumption of a
spherical eye
C. Hernandez-
Matas, X. Zabulis,
A. Triantafyllou, P.
Anyfanti, and A. A.
Argyros
2017
The two images
must not have
bigger difference
and must be
related by a
transformation.
A structure based
region detector for
high-resolution retinal
fundus image
registration
Ghassabi, J.
Shanbehzadeh,
and A.
Mohammadzadeh
2016
The two images
used as input
must have
common
overlapping
regions above a
threshold.
Retinal fundus image
registration via blood
vessel extraction using
binary particle swarm
optimization
P. Palraj and I.
Vennila
2016
The result is not
accurate as
correspondence
at corner points
fails.

SOFTWARE REQUIREMENTS
Coding
Platform
OS
DeviceTool
Matlab
R2013b
Windows
GUIDE
HARDWARE REQUIREMENTS
CPU
Memor
y
Disk
Display
Intel 2.1GHZ
2GB
40GB
15 inch monitor

The modules in the project are
 Graph Extraction
 Optimal Graph Matching
 Fundus Registration
SYSTEM ARCHITECTURE

 Graph extraction module constructs graph for the two images
given as input.
 Optimal graph matching module establishes the
correspondence between the blood vessel graphs and
eliminates noticeable outliers with the help of Optimal Transport
(OT) Algorithm.
 Fundus registration module does image registration using
coefficient algorithm.

DATAFLOW DIAGRAM
The image registration process is split to sub process as shown

The class diagram has the following classes
o Registration class is the user interface class to access the
functionalities of image registration.
o Graph Extraction class constructs graph from the image.
o OT Matcher class finds the correspondence between the two
graphs and filtering the outliers.
o Fundus Registration class implements transformation to fuse
the two images based on the correspondence created by OT
Matcher.

Admin can do the following:
1. Load two fundus images
2. Do registration
3. View output images
4. View performance

1. Pre-processing
The code snippet for pre processing is given:
function im=applyHE(img_gray)
[r c]=size(img_gray);
b=zeros(r,c);
hp_fil=[-1 2 -1;0 0 0;1-2 1];
im=histeq(img_gray);
b=imfilter(img_gray,hp_fil);
c=b+img_gray+25;
medfilt2(c);
im=c;
IMPLEMENTATION

The code snippet for segmentationis given :
T = graythresh(c);
bw = im2bw(c,T+0.3);
bw5=watershed(bw);
SEGMENTATION

FEATURE EXTRACTION
The code snippet for featureextraction is givenbelow :
function fullfeat=extractFeat(im)
[Contrast Energy Homogeneity Mean StandardDeviation
Entropy RMS Smoothness Kurtosis Skewness] =
extractF(im);
I=im;
hsvHist = hsvHistogram(I);
autoCorrelogram = colorAutoCorrelogram(I);
color_moments = colorMoments(I);
img= double(rgb2gray(I))/255;
wavelet_moments = waveletTransform(I);
set = [hsvHist autoCorrelogram color_moments
wavelet_moments]

Classifier Sensitivity(%) Specificity(%) Accuracy(%)
Naïve Bayes 92.90 97.67 92.86
Multi Layer
Perceptron (MLP)
91 93.02 94.05
K Nearest
Neighbors (KNN)
95.12 95.35 95.24
SupportVector
Machine (SVM)
95.89 100 97.39
RESULT AND DISCUSSION

Conclusion
• An automated and flexible technique based on optimal mass transport
theory to tackle the problem of fundus image registration.
• The designed framework yields high-accuracy registrations including many
cases that are difficultto be handled in real circumstances.
Future Enhancement
• As afuture work we will extend the solution for multi modal images.
• This particular change simply consists in providing raw images as input to
the feature extraction step.
• The gradients of both color and grayscale for multimodal images can be
taken instead of the raw images.
CONCLUSION AND FUTURE
ENHANCEMENT

RESULT COMPARISON AND
EVALUATION
o We present an extensive set of comparisons involving our
framework against several state-of-the-art methods
traditionally used in the context of fundus image registration.
o In the following, we first detail the measures and datasets
used to conduct the experiments.

Color labeling-based inspection
The composition of two binary vessel images BI
(green) and BR (magenta) are used to drive the visual
analysis. A large amount of white pixels means better
registration outcomes.

Qualitative analysis: We start our discussion showing how
visual inspections can be properly conducted when
view_x0002_ing the registration results.
• Figure 8 illustrates the process of inspecting the matching
between the fundus images, by counting the amount of
overlapped pixels in the registered images.
• More specifically, we decompose into RGB channels the binary
vessels of the fundus images BI and BR so that the green
component (0,1,0) is taken from BI while its counterpart, the
magenta color (1,0,1), is designated to the binary image BR.
• These cropped components are then overlaid so that the
resulting image will have white pixels(1,1,1) = (0,1,0)+ (1,0,1),
denoting a perfect matching, and green/magenta colors.

1 Danilo Motta, Wallace Casaca, Afonso Paiva, Vessel optimal Transport for
Automated Alignment of Retinal Fundus Images,2019
2 J. Wang, J. Chen, H. Xu, S. Zhang, X. Mei, J. Huang, and J. Ma, “Gaussian field
estimator with manifold regularization for retinal image registration,” Signal Processing,
vol. 157, pp. 225 – 235, 2019.
3 C. Hernandez-Matas, X. Zabulis, A. Triantafyllou, P.Anyfanti, and A. A. Argyros,
“Retinal image registration under the assumption of a spherical eye,” Comput. Med.
Imaging Graph., vol. 55, pp. 95–105, 2017.
3 L. Chen, Y.Lian, Y.Guo, Y.Wang, T. S. Hatsukami, K. Pimentel, N. Balu, and C.
Yuan, “A vascular image registration method based on network structure and circuit
simulation,” BMC Bioinformatics, vol. 18, no. 1, p. 18:229, 2017.
4 Z. Ghassabi, J. Shanbehzadeh, and A. Mohammadzadeh, “A structure based region
detector for high-resolution retinal fundus image registration,” Biomed. Signal Process.
Control, vol. 23, pp. 52–61, 2016.
5 P.Palraj and I. Vennila, “Retinal fundus image registration via blood vessel
extraction using binary particle swarm optimization,” J. Med. Imaging Health, vol. 6,
no. 2, pp. 328–337, 2016.
REFERENCES

deep learn about blood vessel auto1.pptx

deep learn about blood vessel auto1.pptx

Recommended

Recommended

More Related Content

Similar to deep learn about blood vessel auto1.pptx

Similar to deep learn about blood vessel auto1.pptx (20)

More from Priyadarshini648418

More from Priyadarshini648418 (8)

Recently uploaded

Recently uploaded (20)

deep learn about blood vessel auto1.pptx