Professional Documents
Culture Documents
Synopsis on
By
Affiliated to MAKAUT
We are submitting the synopsis on Retinal blood vessel extraction as a part of our final year
seventh semester project under the guidance of Dr. Ranjit Ghosh
Mission –
PEO
PO1 PO2 PO3 PO4 PO5 PO6 PO7 PO8 PO9 PO10 PO11 PO12
Justification :
PSO
Index Page
1. Introduction
1.1 Problem Statement
1.2 Problem Definition
1.3 Objective
1.4 Literature Survey / Background Study
1.5 Brief Discussion on Problem (maximum three pages)
1. Introduction
Among the modern health care community, medical imaging has become the most important tool;
this is because of the visual documentation and record storing for the patients and for its ability to
information extraction about many diseases. For retinal image analysis, there are several
applications, such as diabetic retinopathy, where it can be used in addition to their important roles
in some disease detection in early stages [1, 2]. Because of segmentation of retinal image structures
using in modern ophthalmology as a non-invasive diagnosis; it has been a very interested topic. For
retinal diseases such as hypertension, diabetic retinopathy, hemorrhages, macular degeneration,
glaucoma, neovascularization and vein occlusion, working on the optic disc and the retinal blood
vessel morphology is one of the basic indicators for assessing the presence and severity for each of
these diseases. Assessment of the retinal blood vessels diameter and tortuosity or the optic disc
shape manually has many disadvantages such as, time consuming and prone with human error,
especially with complicated vessel structure and a large number of images [3].
Some of diseases listed above such as glaucoma, diabetic retinopathy, and macular degeneration
are very dangerous. If they aren’t detected correctly and in time, they can lead to blindness.
Therefore, an accurate automated segmentation approach for retinal blood vessel and optic disc is
a very important issue in computer aided diagnosis [2]. The automation of segmentation and
investigation of retinal blood vessel features helps ophthalmologist and eye care specialists to carry
out mass vision screening exams for retinal diseases detection in early stages and treatment
evaluation. This could help at prevention and reduction of vision impairments; age related diseases
and many cardiovascular diseases as well as screening cost reduction [4].
1.3. Objective
This paper proposes a technique to extract blood vessels from a retinal fundus image. From the
input fundus image, the model must be able to extract all the blood vessels using segmentation
and morphological operations. Proposed method is divided into three parts. First, we take the
input and preprocess the input image. Then, the preprocessed image is used to extract the blood
vessels using segmentation algorithm. After this, we use morphological operations as the
postprocessing step to remove the noise and false blood vessels.
Another method is morphological processing which consist of techniques [7] dealing with
digital image processing using mathematical morphology by applying some structure element
(SE) to binary images and sometimes to gray-level images. Roychowdhury, Koozekanani and
Parhi developed a novel three-stage blood vessel segmentation algorithm. The first stage is pre-
processing by high-pass filtering then extracting a binary image and another binary image is
reconstructed from morphologically enhanced image for the vessel regions. Next the major
vessels are extracted which is common regions from these two images. Then the second stage,
Gaussian Mixture Model (GMM) classifier is used to classify all pixels in the two binary images
which are remained from previous stage. Set of 8 features are used in GMM which extracted
depending on first and second order gradient images and pixel neighbourhood. Finally, the last
stage, both the vessels and the classified vessel pixels are combined. This method with accuracy
of 0.952 on the DRIVE.
Marin [8] performed a neural network scheme for pixel classification; 7-D vector was
computed. For training and classification, they used multi-layer feed forward neural network.
There are five main layers, the first one is the input layer which consists of seven neurons, and
the second three layers are the hidden layers consisting of fifteen neurons, finally the output
layer which has one neuron. This method has average accuracy, sensitivity and specificity on
the DRIVE database 0.9452, 0.7067 and 0.9801, respectively.
Morphological multi-scale enhancement method is also presented in [9]. For the extraction of
the blood vessels in the angiogram; fuzzy filter and watershed transformation are used. Multi-
scale non-linear morphology opening operators with structuring element which vary in size is
used to estimate the background, and then the background is subtracted from the image to
In [10] an approach is presented which is combined unique vessel centerline detection with
morphological bit plane slicing. The first order derivative of a Gaussian filter is used in four
directions to extract the centerlines, and then performing an average derivative and derivative
signs with the extracted centerlines. Mathematical morphology has proven their worth as a
brilliant technique for the blood vessels segmentation in the retina. Morphological
multidirectional top-hat operation is applied on blood vessels gray-scale image with linear
structure element to obtain the orientation map and shape, and then the enhanced vessels are
subject to bit plane slicing. For obtaining the vessel tree, these maps are combined with the
centerlines. This method has average accuracy, sensitivity and specificity on the DRIVE
database 0.943, 0.715, 0.977, respectively.
An automated enhancement and segmentation method for blood vessels is presented in [12].
This method decreases the optic disc influence and emphasizes the vessels by applying a
morphological multidirectional top-hat transform with rotating structuring elements to the
background of the retinal image. For producing a vessel response image and the final blood
vessel tree, an improved multi-scale line detector is applied. As line detectors in the multiscale
detector have different line responses, the longer line detectors produce more vessel responses
than the shorter line detectors. To set different weights for different scales, all the responses are
combined by the improved multi-scale detector at different scales. This method has average
accuracy, sensitivity and specificity on the DRIVE database 0.942, 0.735, 0.969, respectively.
Most of these approaches have their advantages and disadvantages, which distinguish it from
other. A complete extraction of the vascular tree in the digital retinal image can be provided
since all the possible vessel pixels have been searched across the whole image. However special
hardware is required to be more suitable for large image dataset for these techniques which are
computationally expensive. In some approaches a significant degradation in the performance is
caused because of noise and lesions so to pick up this noise we perform the enhancement
operation. In our approach we focus on the preprocessing enhancement stage which has a
perfect effect on segmentation step. We use morphological processing to enhance the digital
retinal images to get more accurate results in segmentation process.
Given below is the flow diagram of the model that we have proposed. These are the
sequential steps that we have included to build the model that will extract blood vessel
features from the retinal fundus image.
In this section of the report we have documented the fundamental concepts regarding the different
steps involved in this model. We have provided the different requirements for this this project be
it mathematical, algorithmic, hardware or software. Furthermore, we have presented the
performance with diagrammatic representation during the different stages of the model.
2.1.1. Datasets, comprising both raw fundus images and hand-drawn retinal blood
vessel output image must be created. Different datasets have been used for the
purpose of running tests to determine the values for the thresholding algorithm
and to determine the effectiveness of this proposed method. Here we have used
the DRIVE (Digital Retinal Images for Vessel Extraction) [17] database. It is
available and public database, containing 40 color funds digital images about
population having subjects ranging between 31 and 86 years old. These digital
images were taken by a Canon CR5 non-mydriatic 3CCD camera at 45 field of
view (FOV). Each image is 768 * 584 pixels. The database contains [7] two
main sets, test and training both of each containing 20 digital images. Three
observers, the first and second author and a computer science student manually
figure), resulting in an effective clip limit that is larger than the prescribed limit
and the exact value of which depends on the image. If this is undesirable, the
redistribution procedure can be repeated recursively until the excess is
negligible.
2.1.3. Median Filter, the median filter is a non-linear digital filtering technique,
often used to eliminate noise. This noise reduction is a distinct pretreatment
footstep to upturn the properties of future treatments. The median filter is cast-
off to eradicate salt and pepper noise, to smooth the image. Neighbourhood
averaging can suppress isolated out-of-range noise, but the side effect is that it
also blurs sudden changes such as line features, sharp edges, and other image
details all corresponding to high spatial frequency. The median filter is an
effective method that can, to some extent, distinguish out-of-range isolated
noise from legitimate image features such as edges and lines.
Specifically, the median filter replaces a pixel by the median, instead of the
average, of all pixels in a neighborhood where represents a neighbourhood
defined by the user, centred around location [m,n] in the image.
2.2. Experiments
In this section we have mentioned the different experiments we have conducted in order to
determine the value of the window size and the bias value (C) for the thresholding algorithm.
The window size and the bias value determines how well the retinal fundus image will be
thresholded.
In the first experiment, for selection of C, a series of values from 0.03 to 0.05
with a variation of 0.002 are taken. For every value of C, accuracy is calculated
and plotted against different values of C. This is shown in Fig. 2. From the
graph, it can be observed that the accuracy value increases till the value of
C = 0.042 and then begins to decrease. Therefore, in order to get a better result
with maximum accuracy, the optimum value of C is taken as
0.042. Fig. 4 depicts the output images with different values of C.
From Fig. 3 (a), it can observed that at C = 0.03 the output image includes
vessels with unwanted spurious noise. At C = 0.042 the output image extract
vessels with maximum accuracies 0.9629 as shown in Fig. 3 (b). For higher
value of C = 0.05 less vessels are identified which is shown in Fig. 3 (c).
2.3. Performance
We have not yet performed this step; we are yet to validate the efficiency and accuracy of
this approach. Efficiency or accuracy of the model for each analysis method is analysed
by error rate. This error rate is defined by the terms normal and abnormal right and normal
and abnormal wrong as follows:
1) Abnormal Right (AR): The test gives positive result if extracted vessel is present.
2) Normal Right (NR): The test gives negative result if extracted vessel is absent.
3) Normal Wrong (NW): The test gives positive, but blood vessel is absent.
4) Abnormal Wrong (AW): The test gives negative, but blood vessel is present.
Sensitivity, specificity and accuracy are used to describe the clinical efficiency of the
proposed model.
These are the output of the different stages of the model that we have implemented so far:
A) Raw input fundus image
B) Pre-processed image
C) Thresholded image
D) Post-processed image
Fig. 7. Raw fundus image; Input to the model; stage A (Ordered by: L-R, Top-Bottom)
Fig. 9. Thresholded output containing noise; stage C (Ordered by: L-R, Top-Bottom)
Fig. 10. Final image after applying morphological cleaning operations; Output of the model; stage D
(Ordered by: L-R, Top-Bottom)
This project has been implemented using open source software and libraries, here we have
used Python for the scripting and libraries such as OpenCV, pandas, Scikit, matplotlib etc.
The model can be run in any machine containing Python 3.5 or above interpreter and the
above-mentioned libraries.
Hardware specification:
x64 architecture
Windows 10
8 GB RAM
3. Conclusion
In this paper, we have proposed a method based on local adaptive thresholding for extraction of
retinal vessels. The main contribution of the paper is the application of mean-C thresholding for
retinal blood vessels extraction. This automated method can segment retinal blood vessels
effectively. At the same time, the proposed method is very time efficient as it can complete the
process within a very small execution time. Together the efficacy and strength with its easiness
and fast employment make the offered method a pertinent implementation for being unified into a
comprehensive pre-screening scheme for early identification of ophthalmological disorders.
Although further scope is there to improve the output by tuning the morphological operations for
noise removal of the thresholded output. The performance evaluation is yet to be done.
4. Future Scope
In this project we can improve the blood vessel detection by improving the output. The main
problem in the current state of this project is that the thresholded output produces a very good
blood vessel extract but with a lot of false positives. Thus, by improving the post-processing stage
of the model we can provide a cleaner and more accurate output. The morphological cleaning
operation can further be improved to give a better output.
5. Bibliography
[9] K. Sun, Z. Chen, S. Jiang, and Y. Wang, “Morphological multiscale enhancement, fuzzy filter
and watershed for vascu- lar tree extraction in angiogram,” Journal of medical systems,” in Journal
of medical systems, vol.35, pp. 811-824, (2011).
[10] M.S. Miri and A. Mahloojifar, “An approach to localize the retinal blood vessels using bit
planes and center- line detection,” in Computer methods and programs in biomedicine, vol.108,
pp. 600-616, (2012).
[11] M.M. Fraz et al., “Retinal image analysis using curvelet transform and multistructure
elements morphology by recon- struction,” in Biomedical Engineering, IEEE Transactions on,
vol.58, pp. 1183-1192, (2011).
[12] Y. Hou, “Automatic Segmentation of Retinal Blood Vessels Based on Improved Multiscale
Line Detection,” in Journal of Computing Science and Engineering, vol.8, pp. 119-128, (2014).
[13] Mukhopadhyay, Susanta and Chanda, Bhabatosh, “Multiscale morphological segmentation
of gray-scale images,” Image Processing, IEEE Transactions on, vol. 12, no. 5, pp. ,533- 549
(2003).
[14] M.M. Fraz, P. Remagnino, A. Hoppea, B. Uyyanonvarab, A.R. Rudnicka ,C.G. Owen,
S.A. Barman, “Blood vessel segmentation methodologies in retinal images–A survey, ”
Computer methods and programs in biomedicine, vol. 108, no. 1, pp. 407-433, (2012).
[15] S.Tatiraju,A. Mehta “Image Segmentation using k-means clustering, EM and Normalized
Cuts, ” in Department of EECS, vol. 1, pp. 17, (2008).
[16] Chinki Chandhok,Soni Chaturvedi andA.A Khurshid, “An Approach to Image Segmentation
using K-means Clustering Algorithm, ” in International Journal of Information Tech- nology
(IJIT),vol. 1, (2012).
[17] M. Niemeijer, J.J. Staal, B.v. Ginneken, M. Loog, M.D. Abramoff, DRIVE: digital retinal
images for vessel extrac- tion, http://www.isi.uu.nl/Research/Databases/DRIVE, accessed (2014).