Professional Documents
Culture Documents
Abstract
The basis of study and analysis of various eye diseases includes Optic disc(OD) nerve head region and OD center coordinates.
The early detection of various eye pathologies like glaucoma and Diabetic Retinopathy helps to prevent the vision loss. So, there
is a need to develop a fast and efficient algorithm for disease prediction. For that, reliable and efficient OD localization and
segmentation are the important tasks. Therefore, this method aims for the efficient and automatic localization and segmentation
of Optic Disc from digital fundus images and its peripappilary atropy detection to predict whether the optic disc is affected by
any diseases like glaucoma. The proposed technique is divided into three subsections which deal with OD localization, OD
boundary detection and peripappilary atropy Detection (PPA). OD localization makes use of the unique circular brightness
structure associated with the OD, that is, the OD usually has a circular shape and is brighter than the surrounding pixels whose
intensity becomes darker gradually with their distances from the OD center. OD boundary detection includes accurate blood
vessels inpainting for the removal of vascular structure in the optic disc region which is followed by intensity adjustment and
region growing considering OD center as a seed point for reliable segmentation of fundus images. The presence of PPA indicates
whether the eye is affected by the diseases like glaucoma. So from the segmented optic disc, PPA is detected using a threshold so
that we can predict whether the person is affected by disease or not. This is done by calculating the Red by Green ratio for each
pixel in the Region of Interest (ROI). The process is implemented in a MATLAB 2014 prototype and tested with images in the
DRIVE Dataset. The results show that the suggested method has 82.14% of accuracy, 75% of precision and 81.82% of recall for
segmentation and PPA prediction.
Keywords: Optic Disc, Inpainting, Active Contour segmentation, Ellipse Fitting, Peripappilary Atropy
_______________________________________________________________________________________________________
I. INTRODUCTION
The eye is an organ that reacts to light and has several purposes. The light from an object outside the eye is imaged on the retina.
Optic disk, fovea, and blood vessels are the main components of retina. Out of them, Optic disc (OD) is one of the main
components on retina. It is only part of the retina that is insensitive to light and it is an indicator of various ophthalmic
pathologies. That is, any change in the shape and color of the optic disk is an indicator of various ophthalmic pathologies,
especially for glaucoma which is one of the most common causes of the blindness. Now a days lots of disease can occur in the
eye. Diabetic retinopathy and glaucoma are some of the diseases which cause the loss of vision. These diseases can be examined
by ophthalmologists. Early diagnosis and proper treatment of these diseases can prevent visual loss. Due to population growth
and aging this process requires more number of ophthalmologists. So, there is a need of automatic recognition system for optic
disc detection to identify these diseases in early stage. Optic Disc detection is an important step to examine severity of some
diseases such as glaucoma and diabetic retinopathy. In the computer-aided process of detection and screening of the disease
using color fundus images, automated localizaion of the optic disk and estimation of its boundary are the first two essential steps
before any further analysis. Moreover, the optic disk boundary acts as a reference contour to measure other disk parameters, such
as cup edges. However, it is not a simple task to segment the optic disk reliably and accurately, as there are problems of blood
vessel occlusions, ill-defined edges, noises and fuzzy contour shapes. These problems to a large extent are very similar to other
boundary detection and image segmentation problems in medical imaging area that still require robust solution.
In recent years, several image processing algorithms have been developed to that extent, focusing on analysis of optic nerve
head area in general and optic disc detection in particular. Some features of optic disc like cup-to-disc-ratio, neuro retinal rim,
blood vessels in the different regions of the optic nerve head are used to classify the images as glaucomatous or normal.
Therefore optic disc localization and segmentation forms the significant preliminary step in most of the automatic -computer
assisted diagnosis (CAD) of eye diseases.
The segmentation of the optic disc is a seriously challenging issue. The existing approach performs OD localization and
segmentation efficiently. This is done by blood vessel removal using a strategic inpainting based method for accurate
segmentation of Optic disc using Active contour based region growing approach. The initial step of this approach is to localize
the optic disc using the concept that OD is brighter and center is detected by calculating the centroid of the region of interest
which is used as the initial seed for the region growing approach. But the main problem of this approach is that it cannot
determine whether the optic disc is affected by peripappilary atropy which indicate many eye diseases like glaucoma and diabetic
retinopathy.
To overcome such limitations a new approach has been developed which performs the optic disc localization and
segmentation with peripapillary atrophy detection. This method makes use of all the contribution from the existing approach with
additional functions. That is, After OD center is detected, blood vessels have been successfully inpainted in the optic nerve head
region. In addition, Intensity adjustment is done in optic disc region, which will help the region to grow in an efficient manner.
The usage of active contour segmentation has been devised and employed for accurate segmentation of OD.
Next we have to detect the peripapillary atrophy(PPA) pixels. For that we make use of a threshold-based algorithm to detect
the PPA that helps in the accurate segmenation. The proposed algorithm calculates the Red by Green ratio for each pixel in the
Region of Interest (ROI) and predict whether the optic disc is affected by disease or not.
For the accurate prediction, the classifier used is Random Classifier(RF). Random forests or random decision forests are an
ensemble learning method for classification, regression and other tasks, that operate by constructing a multitude of decision trees
at training time and outputting the class that is the mode of the classes (classification) or mean prediction (regression) of the
individual trees. The main principle behind ensemble methods is that a group of weak learners can come together to form a
strong learner. Each classifier, individually, is a weak learner, while all the classifiers taken together are a strong learner ".
The random forest combines trees with the notion of an ensemble. Thus, in ensemble terms, the trees are weak learners and the
random forest is a strong learner.
There exists many methods for OD Localization and segmentation. In [2] Sinthanayothin et al. presented a method for OD
localization based on the intensity variation in the optic disc region because of blood vessels. This method recognises
automatically the main components of the digital fundus image such as the optic disc, fovea, and blood vessels. This method has
high accuracy of the detection for the optic disc, blood vessels, and the fovea. It has the capablity to cope with highly variable
images because of the usuage of NN for blood vessel detection. The main problem of this method is that the smaller blood
vessels were more difficult to detect. In [3] M. Lalonde et.al. proposes a new method for optic disc detection using pyramidal
decomposition and Hausdorff-based template matching. It is fast and reliable. That is, the template matching approach has a
good computation time but the technique relies on various assumptions. Moreover, this approach suffers due to the vessel edges
present in and around the optic disc region and a high confidence level is a prerequisite for the reliability of image analysis and
the performance of the approach is highly dependent on the threshold edge map.
In [4] M. Foracchia et.al. propose a new method for detecting optic disc in retinal images by means of a geometrical model of
vessel structure. The method is based on the preliminary detection of the main retinal vessels. This method make use of either set
of blood vessels and hence it can accuratey detect the OD. Also it is possible to detect the OD successfully in highly pathological
images. The problem of this method is that the vessels identified in the image should have sufficient curvature to provide good
sensitivity of the model parameters to the data. Moreover this method highly dependent on the availability of the good portion of
vascular structure in the image, whereas it is independent of the actual visibility (or even presence) of the OD. In [5] J. Xu et.al.
proposes a new method for fully automated detection of optic disk boundary in fundus images. This method is more accurate and
robust to blood vessel occlusions, noises, ill-defined edges and fuzzy contour shapes while maintaining the accuracy. The
proposed method is not sensitive to the reasonable variation of weight ratio and it is robust. The problem is that there can be a the
chance for false estimating the disk center under the disturbance of large bright lesion.
In [6] S. Lu et.al. proposes a new method for automatic optic disc detection from retinal images by a line operator. The
proposed method contains retinal preprocessing,designed line operator (detect circular regions that have similar brightness
structure as the OD), and OD detection. The designed line operator is robust against lower image resolution. The main problem
of this method is that line operator is designed based on the assumption that the OD is more or less brighter than the surrounding
retinal pixels and, therefore, cannot handle a very small number of retinal images whose OD is even darker than the surrounding
pixels. That is, this method cannot handle the retinal images that do not have a clear circular brightness structure around their
OD. In [7] G. Joshi et.al. proposes a novel method for segmenting the Optic Disk and Cup from Monocular Colour Retinal
Images for the assessment of Glaucoma. It is an automatic OD parameterization technique which is based on the segmented OD
and cup regions obtained from monocular retinal images. The advantage of using this method is it is robust against local gradient
variations and better ability to capture concavities of objects. That is, in short this approach yield a better representation of image
regions and make the model robust to the distractions found near the OD boundaries. But the main problem of this method is that
it doesnot show a good performance in the case where boundary regions are occluded by thick blood vessels.
In [8] H. Yu, et al. proposes a new method for fast localization and segmentation of optic disc in retinal images using
directional matched filtering and level sets. This method provides fast,robust and accurate segmentation. It can be used for cup-
to-disk ratio computation in glaucoma analysis. It is suitable for assisting automatic screening for early signs of eye diseases. But
the main problem of this method is the algorithm fails on retinal images in which the OD is darker than the surrounding pixels,
such as when a large, very bright nerve fibers is adjacent to the OD. Moreover advanced DR and bright exudates clusters affect
the accuracy and success of our algorithm. In [9] Sandra Morales et.al. proposes a new method for automatic detection of optic
disc based on PCA and mathematical morphology. This method is fairly reliable since it works properly on databases with a
large degree of variability. The main problem of this method is that it is not robust in presence of exudates. Due to the
complexity of the fundus images, their high number of elements makes a perfect segmentation difficult.
In [10] Suraya Mohammad, et al. proposes a new method for for segmenting the optic disc in retinal images using texture
analysis. This method has better computational simplicity and also addresses the illumination issue of the retina images by using
the illumination invariance texture measurement. This method also prove to be reliable in segmenting the optic disc. But the
problem is that the method prove to be reliable in segmenting the optic disc especially in an image with good contrast around the
optic disc boundary and where the blood vessels are not very thick. Also it has high value of classification error rate. The poor
performance of this method is observed in images with the presence of severe peripapillary atrophy especially when the optic
disc boundary is completely missing. In[11] R. Preethi Rajaiah et.al. proposes a new method for Detecting the Optic Disc
Boundary and Segmenting the optic cup for Prediction of Glaucoma. This method is fast, reliable and efficient for the diagnosis
of glaucoma. This method cannot detect the severity of the diseases.
The objective of this work is to design and develop a new method for automatic and efficient OD segmentation and peripappilary
atropy detection from the segmented image. The proposed method is based on region growing approach. Center of OD is
detected automatically which is considered as seed point to grow region. Blood vessels are inpainted in the optic nerve head
region for accurate OD segmentation. In addition, Intensity adjustment is done in optic disc region after blood vessels inpainting,
which will improve the region growing process. Hence the proposed method of optic disc segmentation from fundus image is
further divided into three subsections which deal with OD localization, OD boundary detection and peripappilary atropy
elimination. OD boundary detection includes blood vessels inpainting followed by intensity adjustment and region growing
considering OD center as a seed point. Fig.1 shows the flow chart of proposed OD localization and segmentation method and
peripappilary atropy detection.
Fig. 1: Flow chart of optic disc localization, segmentation amd peripappilary atropy detection.
Preprocessing
The proposed method of OD boundary segmentation includes inpainting of blood vessels in optic disc region and region growing
using Active contour segmentation. As a preprocessing step, a 8-bit image is created by combining red channel and green
channel images with appropriate weightage [13].
I ( x ; y ) 0 . 75 I r ( x , y ) I g ( x , y ) (1)
Where Ir and Ig are the red and green channel of RGB fundus image, respectively. The idea behind assigning more weightage
to red component is to maintain the variation in the image across the OD boundary and suppress that across the blood vessels
which are usually much stronger in green component than that in red component.
In-painting of Blood Vessels
A circular sub-image is extracted around the optic nerve head region, centered at OD center for further processing resulting in
OD boundary detection and segmentation. Using this concept, efficiency and robustness in OD detection increases as it reduces
the search region and computational complexity. Radius, r to extract the sub-image is calculated automatically depending upon
the size of image.
1 M N (2)
Radius , r
2 8 5
Where M & N denotes number of columns and rows in image, respectively. This sub-image will cover the entire optic disc
area making it suitable for segmentation.
The proposed algorithm is based on the fact that the intensity of an OD pixel is a function of the distance from the optic disc
center as it is indicated in Fig. 3. The gray level values decrease as the distance from the center increases radially outwards. Also,
all the pixels that are equidistant from the optic disc center lie on the circumference of a circle of a specific radius having similar
intensity values. This idea is used to in-paint the blood vessel values with a gray level value which is otherwise masked by the
blood vessels.
Let a set of pixels, Si be at a equal distance, R from the optic disc center where i = 1,2,3. . .q. . .k.
max s min{ s }
1) Calculate 1 min s
2
2) If sq 1
(a) (b)
Fig. 4: In-painting of Optic Nerve head region (b) OD contour detection using Region Growing and Spline Interpolation.
Fig. 5 represents the results of OD boundary detection after inpainting and OD boundary detection after inpainting plus
intensity adjustment. It can be clearly seen in the figure that the detection of OD boundary is more accurate after inpainting plus
intensity adjustment in comparison to only intensity adjustment.
Fig. 5: Results of OD boundary detection after inpainting and OD boundary detection after inpainting plus intensity adjustment.
Ellipse Fitting
The curvature points obtained using region growing method are stored and used for ellipse fitting of OD after segmentation.
Contour points, (xi,yi), obtained after the region growing are stored in an array, where i = 1,2,3,4. . ., ne.
Let distance from optic disk center to each contour point is given by
d i ( x0 xi ) ( y0 yi )
2 2 (3)
Then, the Maximum distance of the point set to the Optic disk center is considered as the Major Axis distance given by
R m max( d i ) i (4)
Minor axis is given by the distance from Optic disk Center to point (X1, Y1) as shown in Fig. 6. Point (X1, Y1) corresponds
to a point which has the least column index of all points.
Fig. 8: Result
The method was implemented in a MATLAB 2014 prototype and tested with images in DRIVE dataset. The image processing
were performed on a desktop PC with the following characteristics: Intel Core i3 CPU, 3.4 GHz, 4 GB RAM.
In order to measure the performance of the proposed method three metrics were used: precision, recall and accuracy. Precision
refres to detection exactness and it is defined by (2); recall is related to the detection completeness and it is described by (3); and
accuracy refers to the average correctness of the process and it is defined by (4).
precision TP TP FP (9)
recall TP TP FN (10)
Accuracy TP TN TP FP TN FN (11)
Where, TP stands for True Positive (correctly detected), FP stands for False Positive (incorrectly detected), TN stands for True
Negative(correctly not detected), FN stands for False Negative (incorrectly not detected).
The Proposed method provides better performance on DRIVE Database. The dataset comprised of certain number of images
with different retinal diseases like diabetic retinopathy, glaucoma. The proposed method achieves 82.14% of accuracy, 75% of
precision and 81.82% of recall. Fig. 9 shows the ROC curve for the classification of proposed method.
In the figure points above the diagonal represent good classification results (better than random), points below the line
represent poor results (worse than random). The TPR defines how many correct positive results occur among all positive samples
available during the test. FPR, on the other hand, defines how many incorrect positive results occur among all negative samples
available during the test.
Fig. 10 presents the results of the OD contour detection using the region growing method. The optic disc, its in painted version
and the region detected by the region growing method is presented in the figure.
An extremely efficient element of our proposed OD segmentation method is the in-painting of blood vessels before OD
segmentation using region growing. To estimate the effect of the blood vessels in-painting technique used, accuracy of the OD
boundary segmentation is compared with blood vessels inpainting and without blood vessels inpainting. Table 1 presents that the
proposed inpainting based is effective to accurate OD segmentation. Proposed blood vessels inpainting method is efficient and
novel approach which improves the performance of proposed OD segmentation procedure.
Table 1
Comparison of OD Contour detection with and without Blood vessels inpainting
Finally, the method aims in detecting whether the Optic disc is affected by disease or not by detecting the presence of
Peripappilary Atropy pixel in the fundus image. Table 2 represents that the proposed Disease detection based on OD
segmentation is effective.
Table 2
Peripappilary Atropy Detection for disease prediction.
V. CONCLUSION
In this paper, a complete framework is presented for fast, fully and automatic localization of OD, its accurate segmentation based
on blood vessel inpainting technique and peripappilary atropy detection for predicting whether the eye is affected by any diseases
like Glaucoma. The blood vessel inpainting, region growing method and ellipse fitting is used to segment the OD. From the OD
segmented image, peripapillary atrophy detection is done by training the database in such a way that the Red by Green ratio for
each pixel in the Region of Interest (ROI) is calculated and finding the threshold range for PPA pixels so that when a test image
is given again the same ratio is calculated and if it is between the aforementioned threshold then the image can be predicted as
disease affected image. The proposed OD segmentation algorithm fared well on DRIVE database with an accuracy of 82.14%.
Future work may aim to design algorithms for classifying various eye disease and to determine its severty
REFERENCES
[1] M. Partha Sarathia, Malay Kishore Dutta, Blood vessel inpainting based technique for efficient localization and segmentation of optic disc in digital
fundus Images, Biomedical Signal Processing and Control 25 (2016) 108117.
[2] C. Sinthanayothin, et al., Automated localisation of the optic disc, fovea, and retinal blood vessels from digital colour fundus images, Br. J. Ophthalmol.
83(8) (1999) 902910.
[3] M. Lalonde, M. Beaulieu, L. Gagnon, Fast and robust optic disc detection using pyramidal decomposition and Hausdorff - based template matching, IEE
Trans.Med. Imaging 20 (11) (2001) 11931200.
[4] M. Foracchia, E. Grisan, A. Ruggeri, Detection of optic disc in retinal images by means of a geometrical model of vessel structure, IEEE Trans. Med.
Imaging 23(10) (2004) 11891195.
[5] J. Xu, O. Chutatape, P. Chew, Automated optic disc boundary detection by modified active contour model, IEEE Trans. Biomed. Eng. 54 (3) (2007)
473482.
[6] S. Lu, J.H. Lim, Automatic optic disc detection from retinal images by a line operator, IEEE Trans. Biomed. Eng. 58 (1) (2011) 8894.
[7] G. Joshi, J. Sivaswamy, S. Krishnadas, Optic disc and cup segmentation from monocular color retinal images for glaucoma assessment, IEEE Trans.
Med.Imaging 30 (6) (2011) 11921205.
[8] H. Yu, et al., Fast localization and segmentation of optic disc in retinal images using directional matched filtering and level sets, IEEE Trans. Inf.
Technol.Biomed. 16 (4) (2012) 644657.
[9] Sandra Morales, Valery Naranjo, Jess Angulo, Automatic detection of optic disc based on PCA and mathematical morphology, IEEE Trans.Med.
Imaging 32 (4) (2013) 786796.
[10] Suraya Mohammad, et al., Texture analysis for the segmentation of optic disc in retinal images, in: IEEE International Conference on Systems, Man, and
Cybernetics (SMC), 2013, pp. 42654270.
[11] R. Preethi Rajaiah, R. John Britto, Optic Disc Boundary Detection and Cup Segmentation for Prediction of Glaucoma, in Proc. IEEE ICCV, Vol 3, Issue
10, October 2014.
[12] Jahin Majumdar, A Threshold based Algorithm to Detect Peripapillary Atrophy for Glaucoma Diagnosis, International Journal of Computer Applications,
Volume 126 No.12, September 2015.
[13] Lu Shijian, Accurate and e_cient optic disc detection and segmentation by a circular transformation, IEEE Trans. Med. Imaging 30 (Dec) (2011)
21262133.
[14] M. Bertalmo, et al., Image Inpainting, in SIGGRAPH, ACM Press/Addison-Wesley Publishing Co, New Orleans, USA, 2000, pp. 417424.
[15] A. Criminisi, P. Prez, K. Toyama, Object removal by exemplar based inpainting,in: Conf. on Computer Vision and Pattern Recognition, vol. 2, 2003,
pp.721728,IEEE.