Retinal Image Analysis For Diabetic Patients

Published Date: 02 Nov 2017

Jofin Lal Joe.J

[email protected]

Abstract -Retinal images play important roles in finding of some diseases in early stages, such as diabetes, which can be performed by comparison of the states of retinal blood vessels. Automated image processing has the potential to support in the early detection of diabetes, by detecting changes in blood vessel diameter and patterns in the retina. Intrinsic characteristics of retinal images make the blood vessel detection process difficult. Here, we proposed a new algorithm to detect the retinal blood vessels effectively by using the curvelet transform in representing the edges. The directionality feature of the multi structure elements method makes it an effective tool in edge detection. Afterwards, morphological operators by reconstruction eliminate the ridges not belonging to the vessel tree while trying to preserve the thin vessels unchanged. There is a deficiency of missing some thin vessels is because of utilizing a simple thresholding method. So in this paper I am  replacing the simple threshold method with a adaptive  threshold method  approach to generate the threshold automatically  to increase the accuracy  and reduce the  problem of presence of severe lesions in retinal images.

Keywords—Blood vessel segmentation, curvelet transform, morphology operators by reconstruction multistructure elements morphology, retinal image.

I.INTRODUCTION

Diabetes is a disease that affects about 5.5% of the population worldwide, a number that can be expected to increase significantly in the coming years. About 10% of all diabetic patients have diabetic retinopathy, which is the primary cause of blindness in the Western World. Since this type of blindness can be prevented with treatment at an early stage, the WHO advises yearly ocular screening of patients. The morphological changes of the retinal blood vessels in retinal images are important indicators for diseases like diabetes, hypertension and glaucoma. Thus the accurate segmentation of blood vessel is of diagnostic value Automation will facilitate this screening .Knowledge about the location of the vessels can aid in screening of diabetic retinopathy. Diabetic retinopathy is asymptomatic in early stages of the disease As the disease progresses symptoms may include Blurred vision ,Fluctuating vision,Distorted vision ,Dark areas in the vision ,Poor night vision,Impaired color vision,Partial or total loss of vision

Vessels, fovea, and optical disk are the three most important structures of the human retina and are mostly used for several applications.Detection of these important structures manually is time consuming and depends on the expertise of the user. The segmentation of blood vessels from fundus photographs can be difficult for a number of reasons. Some of the corrupting sources are related to the acquisition process and kind of imagery, and others are intrinsic features of retinal images. The two most influential factors that make the segmentation difficult are the improper retinal image contrast and the uneven background illumination.In other words, arteries have higher contrast than veins. Existing papers also have a deficiency of missing some thin blood vessels because of the simple thresholding method.

Many efforts have been made and various methods have been introduced in order to segment retinal images. The algorithms in this field fall in three groups: window-based, classifier-based and tracking-based approaches. The Window-based methods , such as edge detection, estimate a match at each pixel for a given model against the pixel’s surrounding window. The cross section of a vessel in a retinal image was modeled by a Gaussian-shaped curve in and then detected using rotated matched filters. Classifier-based methods perform in two stages. First, a low-level algorithm produces a segmentation of spatially connected regions. These candidate regions are then classified as being vessel or not vessel. The method proposed in is based on fuzzy K-median clustering, where the connected regions are detected by applying 12 rotating 16 × 15 matched filters, and the results go into a classifier. The final result is produced by length filtering. Tracking-based methods utilize a profile model to incrementally step along and segment a vessel. In order to start tracking, there is a need for seed points. Generally, there are two approaches to select the seed points: manually selecting seeds, which is labor intensive and depends on the expertise of the user and automatically selecting seeds. Vessel segments, which are shorter than a given threshold or shorter than 30 pixels and with a height-to-width ratio bigger than a given threshold,are removed.

In this paper a method based on using curvelet transform is proposed to enhance and prepare the retinal image for better vessel detection. Curvelet as a geometrical transform has two important features: anisotropy scaling law and the directionality. These two features made curvelet capable of sparse representation and handling image singularities better than other multiscale transforms. The second generation of curvelet transform, which is faster and simpler than the first version. Therefore, we used the second generation of curvelet transform, discrete curvelet transform (DCT), and modified the DCT coefficients by a suitable nonlinear function. Oneway to increase the image contrast is to enhance the image ridges, which play an important role in enhancing image contrast.In order to simultaneously enhance the weak edges and eliminate the noise, the modifying function parameters are defined based on some statistic features of fast DCT (FDCT) coefficients. The directionality feature of the multistructure elements method makes it an effective tool in edge detection. Therefore, in the following step, mathematical morphology using multistructure elements are applied to obtain the image ridges. Then, morphological opening by reconstruction helps to remove the detected ridges not belonging to the vessel tree while preserving the thin vessel edges. The morphological opening by reconstruction benefits from using multistructue elements, which helps to improve the performance of this step. There is a restriction on size of structure elements (SEs) concerning the blood vessels diameter. Therefore, the remaining false edges will be removed by means of connected components analysis (CCA) along with length filtering. In order to act locally, image is decomposed to several tiles and CCA, and length filtering is individually applied to each tile.

II.Curvelet Transform

In this paper Curvelet Transform is used in order to propose to enhance and prepare the retinal image for better vessel detection. Curvelet as a geometrical transform has two important features: anisotropy scaling law and the directionality . these two features made curvelet capable of sparse representation and handling image singularities better than other multiscale transforms. Curvelet Transform is a new multiscale transform, curvelet transform is used to overcome the existing drawback of the classical multiresolution approaches such as wavelets. It can represent the edges along curves much more efficiently than the traditional wavelet

A.Discrete Curvelet Transform (DCT)

There are two approaches to implement the so-called second generation DCT: Wrapping method and Unequispaced Fast Fourier Transform (USFFT) method. The wrapping method is faster and easier to implement than the USFFT method. Hence, wrapping method is used in this paper.The FDCT via wrapping: first and unlike earlier discrete transforms, this implementation is a numerical isometry; second, its effective computational complexity is 6 to 10 times that of an FFT operating on an array of the same size, making it ideal for deployment in large scale scientific applications. The wrapping method assumes a regular rectangular grid to wrap the object. The idea is to first decompose the image into a set of frequency bands and to analyze each band by a curvelet transform. The block size can be changed at each scale level.

III. BASIC MATHEMATICAL MORPHOLOGICAL THEORIES

A.Theory

Mathematical morphology is a powerful tool in dealing with various problems in image processing and computer vision. Mathematical morphology is composed of a series of morphological algebraic arithmetic operators. The shape and the size of SE play crucial roles in such type of processing and are, therefore, selected according to the need and purpose of the associated application.

The edges of an image can be found by applying a morphological edge detector named the top-hat transformation described as follows: top-hat(I) = I − (I ◦ S) where (◦) denotes the opening operator. There is a problem in utilizing the top-hat, because pixels in opened image have smaller or equal gray-level values than those in the original image; therefore, the result of top-hat operator includes all the small ordinary intensity fluctuations that can be found in the data such as noise. In addition, the uneven background illumination of the fundus images aggravates this problem severely. To overcome this drawback, a modification was proposed. In the modified top-hat, a closing operator that proceeds by an opening is applied to the original image; the result will be compared to the original image using a minimum operator to attain an image equal to original image except in edges. The modified top-hat transformation is represented as follows

top-hat(I) = I − min((I • Sc ) ◦ So; I) where Sc and So stand for the SEs for closing (•) and opening (◦) operators, respectively.

Opening and closing are two important operators from mathematical morphology. They are both derived from fundamental operations of erosion dilations. Like those operators they are normally applied to binary images, although there are also gray level versions. The basic effect of an opening is somewhat like erosion in that it tends to remove some of the foreground pixels from the edges of regions of foreground pixels. However it is less destructive than erosion in general.

As with other morphological operators, the exact operation is determined by a structuring element. The effect of the operator is to preserve foreground regions that have a similar shape to this structuring element, or that can completely contain the structuring element,while eliminating all other regions of foreground pixels.

The opening of A by B, denoted as AO B, is given by the erosion by B, followed by the dilation by B, that is

AOB = (A Ө B) ⊕ B

Closing is similar in some ways to dilation in that it tends to enlarge the boundaries of foreground regions in an image (and shrink background color holes in such regions), but it less destructive of the original boundary shape. As with other morphological operators, the exact operation is determine by a structural element. The effect of the operator is to preserve background regions that have a similar shape to this structuring element, or that can completely contain the structuring element, while eliminating all other regions of background pixels. Closing is the dual operation of opening and it is denoted by A ● B. It is produced by the dilation of A by B, followed by the erosion by B.

A●B = (A ⊕ B) Ө B

B.Multistructure Elements Morphology

The Segmentation Element(SE) choosing is a key factor in morphological image processing. Single and symmetrical SEs are normally selected in order to perform the morphological processing; such SEs are successful in detecting ordinary, simple and straight edges of an image. The basis of the multistructure elements morphology theory relies on gathering several SEs in one square window

C. Morphological Operators by Reconstruction

The operators morphology closing and opening leave the features larger than SE unchanged. However, the main drawback of conventional opening and closing is that they do not preserve edge information perfectly Basic morphological operators by reconstruction are erosion, dilation, opening and closing.The geodesic dilation of the marker image is defined as the pointwise minimum between the mask image and the elementary dilation of the marker image. The geodesic erosion which is the dual transformation of geodesic dilation is defined as the pointwise maximum between the mask image and the elementary erosion of the marker image.

The morphological opening by reconstruction in its first step eliminates bright features smaller than the SE. in the next step, it dilates iteratively to restore the contours of components that have not been completely removed by opening and it is performed by considering the original image as the reference. In a similar manner, closing by reconstruction is accomplished in case of dark features. Therefore, as a valuable result, producing new edges, edge drift and deforming the contours and edges, which often occur by applying conventional morphological opening and closing will not appear by applying opening and closing by reconstruction

IV.PROPOSED METHOD

In this section, the proposed method is illustrated and the algorithm is described in detail.

A.Input Image Representation Selection

Since the blood vessels in the green channel image of the original colored retinal image have the highest contrast with the background, this channel is chosen to apply the proposed algorithm. The blue channel tends to be empty and the red channel tends to be saturated. The green channel image is suitable for images of DRIVE database.

C:\Users\johnson\Desktop\output screen shot\jof 1.jpgC:\Users\johnson\Desktop\output screen shot\GRN CHN.jpg

Fig.1.a)Input Retinal Image b)Green Channel Image

B.Retinal Image Contrast Enhancement Using FDCT

The curvelet transform is well adapted to represent the images containing edges. It is a good candidate for edge enhancement. A nonlinear function to modify the representation coefficients is introduced in such a way that details of the small amplitude are enlarged. Definition of the function parameters based on some statistical features of curvelet coefficients of the input image is very beneficial to adapt the function better with every input image. The method to enhance the retinal image consists of following steps.

Applying FDCT via wrapping method, a set of scales Sj is obtained, each scale consists of a set of directional bands Di containing coefficients.

For each directional band in each scale Dji, do the following:

Calculate the noise standard deviation σij;

Determine the value of m.

Multiply each coefficient individually by corresponding y.

Reconstruct the enhanced image using modified curvelet coefficients.

The aim of enhancement step is enhancing the thin vessels having low contrast to detect better in the edge detection step. An improper contrast enhancement magnifies the unevenness of background illumination, which causes some false edges in the edge detection step. In other words, the enhancement function enhances all weak edges in the image, which means that the edges of thin vessel and the weak edges arising from uneven background illumination are enhanced coincidently. To overcome this problem, an estimated image background will be subtracted from the enhanced image to decrease the roughness.

C:\Users\johnson\Desktop\output screen shot\GRN CHN.jpgC:\Users\johnson\Desktop\output screen shot\adapative HE.jpg

Fig.2, a)Green Channel Image,b) Adaptive Histogram Equalised image

C. Edge Detection Using Multistructure Elements

In edge detection using multistructure elements morphology, the earlier Structuring Element (SE) of morphological edge detector should be replaced by new introduced SE and follow the following algorithm.

Produce the proposed SE Si with regard to the required directional resolution.

Apply the selected edge detector function F on the original image using the achieved SE in (1) and get the sub edge image F(I)i.

Put the F(I)i obtained in (2) in the following equation to achieve the whole of detected edges.

C:\Users\johnson\Desktop\output screen shot\adapative HE.jpgC:\Users\johnson\Desktop\output screen shot\ADAPTIVE RECON.jpg

Fig.3 ,a)Adaptive Histogram Equalised image,b) AdaptiveReconstructed Image

A simple method to eliminate these undesired objects is achieved by using morphological opening. Morphological opening, besides removing the undesired objects, cause to remove some parts of the blood vessel edges, specifically the thin vessel edges. To overcome this problem morphological opening by reconstruction is used. In order to improve the performance of the morphological opening by reconstruction, the opening using multistructure elements is preferred. Since the multistructure elements are highly sensitive to edges in all directions, it helps to accurately eliminate the false edges.

The SE used in this step is the same as in the edge detection step. The only difference is in assigned weight. Here, instead of assigning weights to each F(I)i , the maximum F(I)i is selected to construct the F(I). This method allows us to eliminate the weak false edges and prevent them from participating in construction of F(I). Therefore, some of undesired objects remain inevitable, which will be removed in length filtering step.

Retina Input Image

Otsu Thresholding (Adaptive Histogram Equalization)

Curvelet in FT Domain

FDCT

Wedge Wrapping

IFCT

IFCT

X

Otsu Thresholding

Length Filtering

Threshold

Choosing Dimension

Morphological Opening

Extracted Blood Vessel

Fig 4)System Architecture

D.Length Filtering

Length filtering is used with the aim of removing the small pixel blocks. The concept of Connected Components Analysis (CCA) is used where connected components pixels which are identified above a specific threshold and labeled using eight connected neighborhood and are considered as a single object, for poor contrast images alpha should be less than one. The large range of gray levels may cause that considering a single threshold for the entire image lead to loss of some parts of thin vessels. In order to deal with this problem, a kind of adaptive CCA, meaning that we consider images in separate tiles and apply CCA and length filtering to each tile individually. By this means, there is no large range of gray levels in each block, and a adaptive  threshold method  approach to generate the threshold automatically  to increase the accuracy  and reduce the  problem of presence of severe lesions in retinal images , i.e Modified CCA is used to predict the length of the blood vessels dynamically here the threshold value is automatically calculated , can be chosen which separates the false edges from vessel edges efficiently. After applying CCA, the components having length less than a specific threshold will be eliminated.

C:\Users\johnson\Desktop\output screen shot\ADAPTIVE RECON.jpgC:\Users\johnson\Desktop\output screen shot\FUSED IMG.pngC:\Users\johnson\Desktop\output screen shot\final image.jpg

Fig.5. a)Adaptive Reconstructed Image .b)fused image.c) Extracted blood vessels

The algorithm is as follows.

Partition image into tiles of N × N pixels with 50% interpolation to avoid windowing effect.

Apply the thresholding algorithm to each part individually and obtain the threshold of each tile.

Apply CCA to each tile with considering only the pixels whose gray levels are more than the corresponding threshold.

Apply length filtering to each tile individually and retain the components having length larger than the corresponding threshold.

Apply opening by reconstruction using multistructure elements to remove the false edges.

In order to eliminate the remained false edges, apply length filtering along with CCA.

Gather all the results in one image.

V. CONCLUSION AND FURTHER WORK

In this paper, a new method for the retinal vessel segmentation has been presented. The retinal image contrast was improved and prepared better for segmentation step , Fast Discrete Curvelet Transformation has been used for the retinal vessel segmentation. Regarding the high ability of FDCT in representing images containing edges, using modification of curvelet transform coefficients, the retinal image contrast was improved and prepared better for segmentation step. Due to high sensitivity of multistructure elements to edges in all directions, multistructure elements morphology was capable of detecting the blood vessel edges successfully. By applying the CCA and length filtering, it helped to remove the remained false edges more accurately. The quantitative performance results of both segmentation and enhancement steps show that this method effectively detects the blood vessels with accuracy of above 94% in less than 1 min.

Because of high sensitivity of multistructure elements to edges in all directions, multistructure false edges, while preserved the thin vessel edges perfectly. By applying the modified CCA and length filtering locally, helped to remove the remained false edges more accurately. Modified CCA predicted all the small length blood vessels dynamically. The quantitative performance results of both segmentation and enhancement steps show that our method effectively detects the thin blood vessels . Hence, my future work is to replace the simple threshold method with a more proper approach to increase the accuracy of the method.