Automated Segmentation of Retinal Blood Vessels using Optimized Gabor Filter with Local Entropy Thresholding and Adaptive Histogram Equalization

Automated Segmentation of Retinal Blood Vessels using Optimized Gabor Filter with Local Entropy Thresholding and Adaptive Histogram Equalization

Mr. V. G. N. S. Prasad M.Tech (P.h.D) Department of ECE ASIST, Paritala, A.P

B. Anitha

B.Tech. Student Department of ECE ASIST, Paritala, A.P

S K. Kaleem Ansari B.Tech. Student Department of ECE ASIST, Paritala, A.P

G. Sai Deepthi

B.Tech. Student Department of ECE ASIST, Paritala, A.P

V. Naveen

B.Tech. Student Department of ECE ASIST, Paritala, A.P

S. Manikanta

1. ech. Student Department of ECE ASIST, Paritala, A.P

   Abstract:- Blood vessel in retinal image plays a vital role in medical diagnosis of many diseases. Diabetic retinopathy is one of the diseases which damages the retina and leads to blindness. Segmentation of blood vessels is helpful for ophthalmologists and this paper presents a new automatic method to extract blood vessels with high accuracy. This algorithm is comprised of optimized Gabor filter with local entropy thresholding for vessels segmentation under various normal or abnormal conditions. The frequency and orientation of Gabor filter are tuned to match that of a part of blood vessels to be enhanced in a green channel image. Segmentation of blood vessels pixels are classified by local entropy thresholding technique in this method. The performance of the proposed algorithm is evaluated by MATLAB software with DRIVE database.

   Keywords:- Retinal image, Blood vessels, Diabetic retinopathy, Optimized Gabor filter, Local entropy thresholding.

   1. INTRODUCTION

      Diabetic retinopathy (DR) is the result of damage caused by diabetes to the small blood vessels located in the retina. Blood vessels damaged from diabetic retinopathy can cause vision loss. Diabetic retinopathy is a leading cause of adult blindness, and screening can reduce the incidence. Screening just increases the chances that a condition will be avoided, found early, or are able to be cured. It is widely recommended that all persons with diabetes mellitus should be regularly screened for diabetic retinopathy.Many image processing methods proposed for retinal vessels extraction. This literature is based on optimized Gabor filter with local entropy thresholding. Gabor filters have been widely applied to image processing and computer vision application problems such as face recognition and texture segmentation [8]. Optimized Gabor filter methods often produce false positive detections and fail to detect vessel of different widths. Also detection

      process much more complicated when retinal image abnormal condition. This paper has been proposed a much robust and fast method of retinal blood vessels extraction using optimized Gabor filter with local entropy thresholding.

      Fig 1: Flow chart of the proposed methodology for blood vessel segmentation in retinal image

   2. MATERIALS AND METHODS

      DRIVE database [30] is used for this analysis. Every image was capture at 584×565 pixels, 8 bits per color in TIFF format. Blood vessels usually have poor local contrast compare to background. The proposed method uses the following steps: (i) Green channel extraction, (ii) Adaptive Histogram Equalization, (iii) Optimized Gabor filter, (iv) Local Entropy Thersholding (v) Binary conversion. The green channel is inverted before the application of the Gabor Filter transform to it, so that the vessels appear brighter than the background. Fig.1. shows the overall procedure of the proposed method.

      1. Preprocessing

         Preprocessing is applied to eliminate the noises in the fundus image. Regarding the acquisition process, retinal images have often low contrast that cause to hardly detect the blood vessels. This method is to improve the image dynamic range to prepare images for next step, detection the blood vessels, and attain to higher accuracy and precision of segmentation. Concerning our purpose, contrast enhancement, the green channel of colored retinal images is used, because compare to other channels it has the highest contrast [4]. Combining advantages of brightness in red channel decreasing the contrast between the abnormalities and the retinal background; this helps to reduce some responses from abnormalities which do not resemble any blood vessels that would otherwise decrease the performance of blood vessels segmentation methods. Contrast-limited adaptive histogram equalization is used for this analysis that enhancing the contrast of the green channel retinal image.

      2. Optimized Gabor Filter

         Gabor filters have been widely used for multi-directional analysis in image processing. In this algorithm optimized Gabor filter is used for detecting the blood vessel in retinal image. The optimized Gabor Filters are a set of orientation and frequency sensitive band pass filters which have the optimal localization in both the frequency contents of the patterns. Theoptimized Gabor filter kernels are sinusoids modulated.

         =

         = /

         = +

         = +

         Where:

         : Standard deviation of Gaussian in x direction along the filter that determine the bandwidth of the filter.

         : Standard deviation of Gaussian filter that control the orientation selectivity of the filter. y

         : Orientation of the filter, an angle of zero gives a filter

         The optimization Gabor filter kernel (9×7 matrix) is rotated in different rotations with the optimized parameters set as follows:

         [. , ], [. , . ], [. , . ]

         = .

         = .

         = ,

         =

         It is required so that the shapes of the filter are invariant tothe scale. The width of the vessels is found to lie within arange of 2-14 pixels (40-200m). Here, and values maintain false positive rate. always (2) rotation phase in this method. The optimized parameters are to be derived by taking into account of size of the lines structures to be detected. Only six optimized Gabor filters with different orientations (0 to 3600 intervals of sixty degrees) are used to convolve with the preprocessing image. The magnitude of each response is retained and combined to generate the result image.

         Fig 2: Original image with extraction channels

      3. Local Entropy Thresholding

      An image can be viewed as an information source with a probability vector described by its grey-level image histogram, the entropy of the histogram can be used to represent a certain level of information contained in the image. Pun [16] and Kapur et al. [17] had taken this concept to derive entropy thresholding methods. However, their approaches did not take into account the correlation among grey levels. As a result, two different images with an identical image histogram will result in the same threshold value [18]. One way to resolve this problem is to consider the grey-level co-occurrence matrix, which contains the information of grey-level transitions in an image. In the proposed method the grey-level co- occurrence matrix developed by Haralick et al [19] is used to derive the Haralick texture feature for retinal image segmentation. The Haralick texture feature chosen is the entropy of the retinal image. In order to performing the proper extraction of the enhanced segments from the Gabor filter response images, an effective thresholding method is required. Assume that a Gabor filter response image has a size of M * N with L grey levels denoted by G = {0, 1. . . L

      1}. A co-occurrence matrix of an image is an L * L square matrix, denoted by whose elements are specified by the numbers of transitions between all pairsof grey levels in G = {0, 1 . . . L 1} in a particular way.

      0 t L-1

      A	B
      D	C

      Fig 3: Quadrants of co-occurrence matrix

      Let t be a value used to threshold an image. It partitions the co-occurrence matrix into four quadrants, namely, A, B, C and D. We assume that pixels with grey levels above the threshold are assigned to the foreground (corresponding to objects), and those equal to or below the threshold are assigned to the background. Then quadrants A and C correspond to local transitions within background and foreground, respectively, whereas quadrants B and D are joint quadrants which represent joint transitions across boundaries between background and foreground. The probabilities associated with each quadrant are then given by

      =

      Obviously 0<pij< 1

      Fig 4: Scatter plot obtained by plotting the local entropy of the optimized Gabor Filter retinal response image

      The entropy threshold determines the optimal threshold t* by maximum of the entropy curve. t* is used as the threshold for segmentation of the retinal image. This Threshold find it automatically form the Entropy- Threshold Curve.

      Fig 5: Local entropy curve of the foreground object form quadrants A and C of the scatter plot

   3. IMAGE ENHANCEMENT AND FEATURE EXTRACTION

      Several abnormal retinal images from four different categories are collected from the hospitals and used in this work for disease identification. The abnormal retinal images collected from the hospitals cannot be directly classified by the automation techniques. The reason is twofold: (a) Lack of clarity in the anatomical features which is mainly due to the poor contrast of the raw image and (b) Large dimensions of the input image which accounts for the complexity of the system. Hence, suitable techniques must be adopted prior to the image classification process to overcome these drawbacks.

      3.2.1 RETINAL IMAGE DATABASE

      The image dataset consists of 420 digital retinal images obtained using the imaging camera. All the images are collected from Lotus Eye Care Hospital, Coimbatore, India. The images are stored as color TIFF images and are 1504×3100 pixels for all the objects. The color image is made up of three channel images (red, blue and green). The intensity value of each channel images range from 0 to 255.

      3.3 Retinal image enhancement

      The enhancement step is a mandatory task in automated image classification system. Retinal images are routinely acquired and assessed to provide diagnostic evidence for many important diseases namely diabetic retinopathy, hypertension etc. Because of the acquisition process, very often these images show important lighting variations, poor contrast and noise. This problem may seriously affect the diagnostic process and its outcome. Pre-processing algorithms are implemented to enhance the original image so that it can increase the chances for success of subsequent steps.

      Retinal images usually have pathological noise and various texture backgrounds, which may cause difficulties in extraction. The raw retinal images usually have very low contrast which is signified by the grouping of large peaks in a small area on the histogram plot. In order to reduce these imperfections and generate images more suitable for extracting the features demanded in the classification step, a pre-processing procedure comprising of six stages is

      applied in this work. The six stages are: (1) Green band processing, (2) Grey conversion, (3) vessel central light reflex removal, (4) Adaptive histogram Equalisation, (5) 2- D optimized gabor filtering .

      3.3.1 Image processing result

      The experiments are conducted on the input 420 images using MATLAB software. Initially, the green channel, blue channel and red channel images are extracted from the raw image. These results are shown in Figure 3.4. A qualitative analysis of these images show the superior nature of the green channel image. Sample pre-processed output images (one from each category) are displayed in Figures 3.5, 3.6,

      3.7 and 3.8.

      (a)

      (b) (c)

      (d)

      Figure 6: Sample channel images. (a) Input image, (b) Red image, (c) Blue image and (d) Green image

      A careful observation of the above channel images has clearly revealed the superior nature of the green channel image. The blood vessels are displayed with much clarity in the green channel image than the other channel images. Thus, the green channel images are extracted from all the images and used for further analysis.

      Fig 7: Green channel of the original image (left) and Adaptive histogram equalization image (right)

      Fig 8: Gabor Filter Response Image

   4. RESULTS AND DISCUSSIONS

      This paper, the software selected to perform the experiment is MATLAB 7.10 (R2010a). The retinal images were collected from the DRIVE database. The accuracy (Acc) is calculated by the ratio of the number of correctly classified pixels to the total number of pixels in the image. This method average accuracy is 97.94%. The sensitivity (Se) represents the fraction of pixels correctly classified as vessel pixels, where the false positive defines the fraction of pixels erroneously classified as vessel pixels. Average sensitivity is 98.5%. The computational time of whole process of our method takes approximate 2 seconds for each retinal image. Accuracy (Acc).

      Accuracy (Acc)

      Acc=TP+TN/TP+TN+FP+FN

      sensitivity (Se)

      Se= TP/TP+TN

      Fig 9: Proposed Segmented Image (left) and Gold Standard Image (right)

      Fig 10 : Compared Image (Gold standard & Final resultant image)

      [True Positive (TP) = Blue pixels, False Positive (FP) = Black pixels,

      False Negative (FN) = Red pixels, True Negative (TN) = White pixels]

   5. CONCLUSION

      A fully automated retinal blood vessels segmentation method has been presented. In this paper, first introduce optimized Gabor filter with local entropy thresholding for vessels extraction. The threshold value obtained by calculating the entropy feature from the gray level co- occurrence matrix of the optimal Gabor filter retinal image can be one of the ways of segmentation of retinal images. This analysis demonstrated higher true positive rate and reduce false detection in retinal image. The performance of the proposed method is evaluated by comparing DRIVE database images.

   6. REFERENCES

1. American Diabetes Association. Standards of medical care for patients with diabetes mellitus. Diabetes Care 2000; 23: S32 S42.

2. AlirezaOsareh and BitaShadgar, Retinal Vessel Extraction Using Gabor Filters and Support Vector Machines, Advances in Computer Science and Engineering Communications in Computer and Information Science Volume 6, 2009, pp 356-363

3. AlauddinBhuiyan, BaikunthNath, Joselito Chua and RamamohanaraoKotagiri, Blood Vessel Segmentation IEEE International Conference on Vol: 5, Publication Year: 2007,

   Page(s): V – 521 – V 524

4. P.C.Siddalingaswamy,K.Gopalakrishna

   Prabhu,Automatic Segmentation of Blood Vessels in Colour Retinal Images using Spatial Gabor Filter and Multiscale Analysis, 13th International Conference on Biomedical Engineering, IFMBE Proceedings Volume 23, 2009, pp 274-276 Springer

5. Wu, D.; Ming Zhang; Jyh-Charn Liu; Bauman, W., "On the adaptive detection of blood vessels in retinal images," Biomedical Engineering, IEEE Transactions on , vol.53, no.2, pp.341,343, Feb. 2006