Elsevier

Applied Soft Computing

Volume 52, March 2017, Pages 937-951
Applied Soft Computing

Optimized clinical segmentation of retinal blood vessels by using combination of adaptive filtering, fuzzy entropy and skeletonization

https://doi.org/10.1016/j.asoc.2016.09.033Get rights and content

Highlights

  • Occasionally certain areas in the retina can be questionable for physicians which can lead to wrong interpretations for patients.

  • A method is proposed that introduces a higher ability of segmentation by employing Skeletonization and a threshold selection based on Fuzzy Entropy.

  • By extracting indices of the human retina properly, physicians will be able to estimate pathological injuries with a higher confidence.

  • The proposed approach is fast and outperforms over other previously competitive techniques.

  • The proposed approach consists of two stages. First of all, the retinal vessels was preprocessed by the HSV space and Wiener Filter. Then, the segmentation level is implemented by using Adaptive Filter that employs optimum threshold based on Fuzzy Entropy and Skeleton algorithm.

Abstract

The analysis of retina blood vessels in clinics indices is one of the most efficient methods employed for diagnosing diseases such as diabetes, hypertension and arthrosclerosis. In this paper, an efficient algorithm is proposed that introduces a higher ability of segmentation by employing Skeletonization and a threshold selection based on Fuzzy Entropy. In the first step, the blurring noises caused by hand shakings during ophthalmoscopy and color photography imageries are removed by a designed Wiener’s filter. Then, in the second step, a basic extraction of the blood vessels from the retina based on an adaptive filtering is obtained. At the last step of the proposed method, an optimal threshold for discriminating main vessels of the retina from other parts of the tissue is achieved by employing fuzzy entropy. Finally, an assessment procedure based on four different measurement techniques in the terms of retinal fundus colors is established and applied to DRIVE and STARE database images. Due to the evaluation comparative results, the proposed extraction of retina blood vessels enables specialists to determine the progression stage of potential diseases, more accurate and in real-time mode.

Introduction

Growing increase of health and therapeutic costs during the past 30 years have forced health and therapeutic researchers as well as authorities to study ways of promoting operational efficiency as well as reducing costs of these burdens. Simulation models are able to show variability, interactions and complexities of a system. Excessive dilation of blood vessels within the retina is recognized as one of the special symptoms of retina disorder. In most studies there is a need to examine the morphological characteristics and antigen expression patterns of cultured human retinal glia to define novel subtypes [1]. Analysis of retina blood vessels has shown to be a very efficient method used for diagnosing diseases such as diabetes, hypertension and arthrosclerosis [2], [3]. Most of the retinopathy of prematurity shows itself by changing characteristics such as thickness, color, torsion or curvature and translucence rate. Among the specialists techniques, excessive dilation of blood vessels of retina − which is more seen in the posterior section of retina – is recognized as one of the special symptoms in retinopathy of prematurity [4]. For this reason, correct analyses of the retinal vessels can lead to more accurate diagnosis, treatment and clinical studies. Inevitably as the study of blood vessels is so difficult, sometimes even expert’s commit errors in their correct diagnoses. For this reason, it is not reasonable to isolate these vessels from the rest of tissue manually. Different techniques have been suggested for accurately observing blood vessels of the retina such as the RetCam system and also using an ophthalmoscope. Researchers have extracted blood vessels of the retina based on the images received from these systems and in the images; the first focus is the isolation of features such as color of vessels, shape and contrast. This strategy is now commonly used as there are sometimes vessels with these features in healthy tissue.

Filters are often used to maximize performance response in the isolation of blood vessels [5], [6], [7], [8], [9]. Other methods which use morphologic operators in image detection are based on mathematical calculations [10], [11]. Some of the methods suggested by researchers rely more on the retina image edge detection technique. Other techniques are also based on the machine learning principles [5], [12], which can be divided into two groups; supervised methods [2], [12], [13] and unsupervised methods [14]. Supervised methods are the ones with which it can be determined if one pixel of the image belongs to blood vessels of the eye retina. The unsupervised methods do the same work but the difference is that no definite classification is done on pixels of the image. Among the traditional methods based on image processing techniques, modeling of cross sections of retina vessels and use of conversion and variation methods based on different filters were first introduced by Chaudhuri et al. [5]. The application of the thresholding method globally or locally [6] and its combination with other filtering methods on images [7] are suitable methods and the only shortcoming is that the selection of thresholding border at the time of isolation is not intelligent. Martinez et al. [15] suggested a method with which some features were received by differentiating the image and it was possible to isolate the area including the pathological vessels.

One of the most important destructions which occur when receiving images from the environment is Motion Image Blurring when frames recorded from video strings blur the image and generally can be in three directions of length, width and depth. In case the object is inherently mobile, this blurring is inevitable. Awareness with creation of this blurring helps us determine how to reconstruct it.

Local discontinuities in brightness of pixels of an image from one level to another level are called image edges [21]. Use of differential detection and model adaptation are of the methods on which basis edge can be detected in the image. Although researchers regard edge detection as one of the efficient techniques for image processing, they have not suggested a suitable and comprehensive solution for accurate detection of edges in images aside from the scale or form of the edge [22]. Images of the retina also require capable techniques which can extract blood vessels from other parts with high precision so that only edges relating to blood vessels can be isolated and other parts which are detected as edges can be deleted from the image. Combination of the edge method, adaptive filtering and growth of working area increases speed of isolation considerably. This method was introduced by Wang et al. [16] and it seemed to be suitable for making the retina vessel isolation process intelligent. Combination of morphological filters [11], extraction of edges of main arteries [2] and classification of pixels of each section in tissue of retina based on a characteristic vector [12] are some of the other techniques researchers have been suggesting for the extraction of blood vessels. In this paper, a new method has been introduced based on the use of an adaptive filter and Gaussian first order derivative function. The main advantage of an adaptive filter along with a differentiator is that other parts of the retina tissue which appear abundantly in the edge detection stage will be deleted by implementing a differentiator on the output of the filter. As a result, it is easier for the specialist to analyze retina vessels. To obtain optimal threshold rate, two parts of the adaptive filter output and its derivative are used and as a result, optimal value of the threshold will be determined based on fuzzy entropy based on membership functions.

More recently, researchers have used different techniques for retinal segmentation. For example, Li et al. [23] used a supervised method with desirable sensitivity and specificity for the segmentation of blood vessels. Also, the method proposed by Krause et al. [24] can be used for rapid analysis of tissue and finding blood vessels using skeletonization technique.

Occasionally certain areas in the retina can be questionable for physicians which can lead to wrong interpretations for patients; where a blood vessel can be labeled as ‘A’ if the vessel is artery, ‘V’ if the vessel is vein, and ‘white squares' if the labeling was questionable [2]. Fig. 1 shows the hand-labeling of four example retinal images in the STARE database which according to samples shows some questionable parts of retinal vessels. These images were obtained from clinics where the patient stays in the clinic so that the physician can check the images, often leading to a lengthy procedure for patients. The first and most practical benefit of the proposed method is extracting retinal vessels based on a physician’s ability to understand the status of a patient's vessels with higher certainty and in a real-time manner.

Experts of this field commonly make decisions about the current test by applying on patients and they compare these results between the analogous patients with each other. Accordingly, an efficient method based on Adaptive Filter that employs optimum threshold based on Fuzzy Entropy and Skeleton Algorithm is introduced and proposed in this paper. The proposed approach is fast and outperforms over other previously competitive techniques.

The rest of this paper is organized in 6 sections. Section 2 presents the background of proposed method. In Section 3, during the preprocessing phase, the HSV space and Wiener filter are used to compensate for the effects of non-uniform lighting and De-blurring respectively. In Section 4, the retinal vessel segmentation is implemented by two techniques: an adaptive filter is used for identifying blood vessels and the optimal threshold is selected based on the fuzzy entropy method. At the end of Section 4, the skeleton producer is applied on the segmented vessels to examine the connectivity of pixels. Section 5 presents experimental results. Discussion and conclusion are given in Section 6 and Section 7, respectively.

Section snippets

Data analysis

To adjust the extracted output with an ideal source, two databases of STARE [2] and DRIVE [7] for retina images were used and the desired vessels were manually extracted from the images by a specialist. DRIVE data included 40 images in which 12 images indicated tissue injuries (such as diabetes or hypertension). The images have been received with a Canon digital camera, CR5 model as 3CCD (Charge-Coupled-Device) and with a field of vision of 45 ° and a size of 584 × 768 pixels. The images received

Deblurring and Denoising

One of the most important filters that can simultaneously try to remove the blurring and noise from original images is the Wiener filter. This filter is defined by an equation so that when power is reduced in it, the Wiener filter will act as the reverse filter. Wiener filter, which seeks to find an optimal compromise between inverse filtering and denoising, is one of linear estimators that use the concepts of orthogonally properties [26]. The frequency response of this filter can be expressed

Adaptive filtering

Adaptive filters were first used for identifying blood vessels [5] in which the desired filter was applied as Gaussian Function for adapting vessels and identifying them with high accuracy. This filter, commonly known as Gaussian Matched Filter, is used to identify vessels in a way that the profile of cross-section can be stated in terms of Gaussian function [38]. This filter not only detects vessels, but also can be effective in detecting the optic disk and light lesions boundary [39]. If Gσ

Experimental results

The presence of noise or blurring in images of the retina has effect on vessels segmentation and consequently detection of effect which will provide different interpretations of the conditions of the retina. By implementing the proposed algorithm in the first step, noise and blurring are removed from the main image. The main advantage of the algorithm in protection of information includes the retina and the specialist will make a comment on the probability of disease and pathological injury

Discussion

The proposed algorithm has some strengths and weaknesses compared with other segmentation techniques. It can be inferred that most of the proposed methods are not resistant to noise or blurring of the images. For this reason, the first priority of the proposed algorithm over other reliable methods is the ability of the algorithm to preprocess the images for preventing entrance to the output and finally extracting vessels of the retina properly.

In Table 3, Table 4, the statistical comparison of

Conclusion

The software introduced in this paper detects blood vessels of the retina with appropriate accuracy based on analysis of the images received from the retina using an adaptive filter and the fuzzy entropy-based thresholding method. Practical results and final simulations of the algorithm showed that although there is little statistical difference between the proposed and human observer images, it provides higher sensitivity. By deleting the optic disk border and light reflections of the

References (56)

  • A. Glen

    The international classification of retinopathy of prematurity revisited

    Arch. Ophthalmol.

    (2005)
  • S. Chaudhuri et al.

    Detection of blood vessels in retinal images using two-dimensional matched filters

    IEEE Trans. Med. Imaging

    (1989)
  • T. Chanwimaluang et al.

    An efficient algorithm for extraction of anatomical structures in retinal images

  • A. Hoover et al.

    Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response

    IEEE Trans. Med. Imaging

    (2000)
  • A.M. Mendonça et al.

    Segmentation of retinal blood vessels by combining the detection of centerlines and morphological reconstruction

    IEEE Trans. Med. Imaging

    (2006)
  • F. Zana et al.

    Segmentation of vessel-like patterns using mathematical morphology and curvature evaluation

    IEEE Trans. Image Process.

    (2001)
  • M. Niemeijer et al.

    Comparative study of retinal vessel segmentation methods on a new publicly available database

    SPIE Med. Imag.

    (2004)
  • S. Garg et al.

    Unsupervised curvature-based retinal vessel segmentation

  • C. Sinthanayothin et al.

    Automated localisation of the optic disc, fovea, and retinal blood vessels from digital colour fundus images

    Br. J. Ophthalmol.

    (1999)
  • J.J. Wang et al.

    Retinal vessel diameters and obesity: a population-based study in older persons

    Obesity (Silver Spring)

    (2006)
  • S.J. Lee et al.

    Costs of mobile screening for diabetic retinopathy: a practical framework for rural populations

    Aust. J. Rural Health

    (2001)
  • S. Kindermann et al.

    De-blurring and De-noising of images by nonlocal functional

    Multisc. Model. Simul.

    (2005)
  • A. Beck et al.

    Fast gradient-based algorithms for constrained total variation image De-noising and De-blurring problems

    IEEE Trans. Image Process.

    (2009)
  • R. Gonzalez et al.

    Digital Image Processing

    (2008)
  • W. Pratt

    Digital Image Processing: PIKS Scientific Inside

    (2007)
  • T. Acharya et al.

    Image Processing: Principles and Applications

    (2005)
  • Q. Li et al.

    A cross-modality learning approach for vessel segmentation in retinal images

    IEEE Trans. Med. Imaging

    (2016)
  • M. Krause et al.

    Fast retinal vessel analysis

    J. Real-Time Image Process.

    (2016)
  • Cited by (52)

    • A review of retinal vessel segmentation for fundus image analysis

      2024, Engineering Applications of Artificial Intelligence
    • An efficient registration-based approach for retinal blood vessel segmentation using generalized Pareto and fatigue pdf

      2022, Medical Engineering and Physics
      Citation Excerpt :

      The sensitivity was 99.14 percent, the specificity was 67.34 percent, and the average accuracy was 98.51 percent. Table 5 shows the comparative analysis of the proposed approach with other retinal vessel segmentation approaches; It has been found that the Generalized Pareto pdf and Fatigue pdf based retinal image segmentation technique outperforms existing vascular segmentation approaches using Deep Learning approach [33,46–54], Machine Learning approach [55–68], Unsupervised Learning approach [69–74], Multiscale approach [75–84], Morphological approach [34,85–88] and Matched Filter approaches [10,35,89–93]. The proposed approach of registration of segmented retinal image pair source and target of Generalized Pareto pdf and Fatigue pdf based matched filter outperforms other vessel segmentation approaches such as Deep Learning-based approach, Machine Learning-based approach, Unsupervised learning approach, Morphological approach, Multiscale approach and Matching Learning approach.

    • Retinal blood vessel segmentation using pixel-based feature vector

      2021, Biomedical Signal Processing and Control
      Citation Excerpt :

      To obtain a classifier is used the feature vector and manual labels. Rezaee et al. [41] suggested an efficient method based on Fuzzy Entropy. Firstly, the blurring noises are eliminated with a Wiener filter provided.

    View all citing articles on Scopus
    View full text