A novel method for blood vessel detection from retinal images
© Xu and Luo; licensee BioMed Central Ltd. 2010
Received: 6 November 2009
Accepted: 28 February 2010
Published: 28 February 2010
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.
In this paper, we present a novel method to segment retinal blood vessels to overcome the variations in contrast of large and thin vessels. This method uses adaptive local thresholding to produce a binary image then extract large connected components as large vessels. The residual fragments in the binary image including some thin vessel segments (or pixels), are classified by Support Vector Machine (SVM). The tracking growth is applied to the thin vessel segments to form the whole vascular network.
The proposed algorithm is tested on DRIVE database, and the average sensitivity is over 77% while the average accuracy reaches 93.2%.
In this paper, we distinguish large vessels by adaptive local thresholding for their good contrast. Then identify some thin vessel segments with bad contrast by SVM, which can be lengthened by tracking. This proposed method can avoid heavy computation and manual intervention.
The retina is the only location where blood vessels can be directly captured non-invasively in vivo. Over the past decade, the retinal image analysis has been widely used in medical community for diagnosing and monitoring the progression of diseases [1, 2]. And retinal blood vessels are important structures in retinal images. The information obtained from the examination of retinal blood vessels offers many useful parameters for the diagnosis or evaluation of ocular or systemic diseases. For example, the retinal blood vessel has shown some morphological changes such as diameter, length, branching angles or tortuosity for vascular or nonvascular pathology, such as hypertension, diabetes, cardiovascular diseases . Blood vessels are also used as landmarks for registration of retinal images of a same patient gathered from different sources. Sometimes, retinal blood vessel must be excluded for easy detection of pathological lesions like exudates or microaneurysms. In all cases, proper segmentation of retinal blood vessel is crucial.
Actually, automatic detection of the blood vessels in retinal images is a challenging task. The contrast of retinal image diminishes as distance of a pixel from the center of the retinal image. And the presence of noise, the variability of vessel width, the presence of some pathological lesions, all make the task more difficult.
There are three basic approaches for automated segmentation of blood vessel: thresholding method, tracking method and machine trained classifiers. In the first method, many of different operators are used to enhance the contrast between vessel and background, such as Sobel operators, Laplacian operators, Gaussian filters which model the gray cross-section of a blood vessel . Then the gray threshold is selected to determine the vessel. And this gray threshold is crucial, because small threshold induces more noises and great threshold causes loss of some fine vessels, so adaptive or local threshold is used to different sections of an image.
Vessel tracking is another technique for vessel segmentation, whereby vessel centre locations are automatically sought along the vessel longitudinal axis from a starting point to the ending point . This method may be confused by vessel crossings and bifurcations.
Many kinds of classifiers, such as Bayesian classifier, neural networks, support vector machine, have been employed for improved discrimination between vessel and non vessel. Feature extraction and parameters selection of a classier are critical. All pixels in images are classified into vessel or non-vessel through the classifier [6, 7].
In fact, a single generally acknowledged vessel segmentation algorithm does not exist due to the unique properties of each acquisition technique. Every segmentation method has difficulties when applied alone, a combination of them is presented to detect retinal blood vessel in this paper. This article is organized as follows. Section 2 describes the method of segmentation of blood vessel. Section 3 shows the results. The discussions and conclusions are presented in Section 4.
Due to the acquisition process, retinal images often have a variational gray level contrast. In general, large vessels display good contrast while the thin ones show bad contrast. Thereby pixels attached to large and thin vessels show the different gray level and geometrical correlation with the nearby pixels. So we extract large and thin vessels separately. The proposed method is made up of four fundamental parts, (1) preprocessing, which involves background normalization, image binarization and large vessel extraction, (2) feature extraction of fragments, which are the residual parts of binary retinal image with large vessels excluded, (3) classification of fragments, support vector machine is used to distinguish thin vessel segments from all the fragments, (4) thin vessel growth, based on tracking method.
Image preprocessing and large vessels extraction
The adaptive local thresholds are implemented to the normalized image, and the binary retinal image is shown in Fig.1 (c). It is obvious that the large connected components are the large vessels. So the large connected components, whose area are greater than T (T is the minimum grain size, and equals 100 in this paper), are reserved as the large vessels candidates, shown in Fig.1(d).
The edges of the optic disk are usually mistaken for large vessels. According to the spatial gray properties of the optic disk and large blood vessels, automatic erasing the edges of the optic disk are implemented subsequently. The optic disk is the origin of blood vessels and the brightest region in retinal image. Large vessels are dark objects with two-sides boundary, relative to the background. In the gradient images convoluted with Sobel operators along horizontal and vertical directions, a large vessel always corresponds to a pair of local gradient maximum and minimum on both sides along a profile. And the edge of the optic disk corresponds to a single local maximum or minimum. We locate the optic disk, reserve the pixels between a pair of local gradient maximum and minimum around the optic disk along horizontal or vertical direction and remove all the small objects nearby. Fig.1(e) shows the binary retinal image where the edges of the optic disk are erased automatically.
As mentioned above, a large vessel corresponds to a pair of local gradient maximum and minimum on both sides. So we exclude the pixels outside the pairs of extrema and fill in the inner parts. By doing this, we obtain the large blood vessels, shown in Fig.1(f).
The residual fragments of the binary image are not all noises, shown in Fig.3(a). They should be classified by support vector machine next.
The wavelet and curvelet transform are all multiscale transforms. They are now recognized as useful feature extraction methods to represent image features at different scales. And the wavelet and curvelet transforms exhibit impressive performance in detecting point and line features, respectively. Nevertheless, the wavelet and curvelet modulus are bigger along the image edges.
where x i_ori is the original i th feature (i = 1, 2Λ 12) of x, while x i _maxand x i _minare the maximum and minimum of i th feature. And the x i ranges from -1 to 1 after normalization.
Classification based on support vector machine
Where the constant C > 0 is user defined and determines the trade-off between the maximization of the margin and minimization of the classification error .
where x is the feature vector, y is the corresponding classification result and σ is width of 'rbf'. Here y is 1 as pixel attached to a thin vessel and -1 to non-vessel pixel.
Growth of thin vessel
where subscripts x and y denote the convolution of original gray image with the second order Gaussian derivate along x or y direction. Then the eigenvalues of Hessian matrix and their corresponding eigenvectors can be calculated. The eigenvalues, λ 1 and λ 2, where we take λ 1 ≥ λ 2, measure convexity and concavity in the corresponding eigendirections. For a retinal image, where vessels are darker than the background, the eigenvalues of the Hessian matrix are λ1 >> 0 and λ2 ≈ 0 for vessel pixels .
The endpoints of the thin vessel segments grow along D 1 direction recursively, until the estimated next thin vessel endpoints can't meet the conditions of λ1 >> 0 and λ2≈ 0.
Results and Discussion
This method is evaluated on the publicly available DRIVE database . The DRIVE database contains 40 color retinal images. These 40 images were divided into a training set and a test set. The binary images of manual segmentation and the masks of field of view (FOV) are available for all the images of the two sets. All the images were manually segmented. And those of the test set were segmented twice by two experienced experts. And the 1st expert marked 12.7% of pixels as vessel, against 12.3% for 2nd expert. The performance of the different segmentation method is usually evaluated on the test set using the 1st manual segmentations as ground truth. The 2nd expert reaches the accuracy of 0.9473, which is usually used as references.
We first select a training set to build the SVM classifier. The training set is made up 2000 pixels, extracted in proportional spacing from the 20 residual images excluded large vessels in the training set (100 pixels per image). There are 1000 thin vessel pixels while the others are due to the presence of noise in the images. Classification performance is then evaluated on all pixels of the 20 residual images of the training set. In order to ensure accurate classification for most noise pixels (about >95%), the SVM is used continually until the number of remained pixel is less than 1000. According to the training set, there are about 1500 thin vessel pixels and more than 10000 non-vessel pixels in each residual image. And less than about 400 thin vessel pixels remained after SVM, so the specificity (ratio of the number of true negatives and the number of all negative samples) can't reach 95% if the reserved pixels are more than 1000.
Performance of segmentations blood vessel in test set
Comparison with some different vessel segmentation methods
Average accuracy (standard deviation)
In this paper, we deal with the retinal blood vessel that appears split into two parts, due to the contrast, large and thin vessels. The large vessels are detected by adaptive local thresholding in normalized images. Then the residual fragments including thin vessel segments are classified by SVM. This method avoids heavy computation that applied SVM to each pixel. And the thin vessel can be extracted by iterative linear extrapolation without manual given start points.
- Ian NM, Patricia MH, R'John W: Image registration and subtraction for the visualization of change in diabetic retinopathy screening. Comput Med Imaging Graphics 2006, 30: 139–145. 10.1016/j.compmedimag.2006.01.002View ArticleGoogle Scholar
- Wong LY, U'Rajendra A, Venkatesh YV, Caroline C, Lim CM, Ng EYK: Identification of different stages of diabetic retinopathy using retinal optical images. Inf Sci 2008, 178: 106–121. 10.1016/j.ins.2007.07.020View ArticleGoogle Scholar
- George KM, Pantelis AA, Konstantinos KD, Nikolaos AM, Thierry GZ: Detection of glaucomatous change based on vessel shape analysis. Med Imaging Graphics 2006, 30: 139–145. 10.1016/j.compmedimag.2006.01.002View ArticleGoogle Scholar
- Mohammed AR, Munib Q, Mohammed A: An improved matched filter for blood vessel detection of digital retinal images. Comput Bio Med 2007, 37: 262–267. 10.1016/j.compbiomed.2006.03.003View ArticleGoogle Scholar
- Xu Y, Zhang H, Li H, Hu GS: An improced algorithm for vessel centerline tracking in coronary angiograms. Comput Meth Prog Bio 2007, 88: 131–143. 10.1016/j.cmpb.2007.08.004View ArticleGoogle Scholar
- Joao VBS, Jorge JGL, Roberto MCJ, Herbert FJ, Michael JC: Retianl Vessel Segmentation Using the 2-D Gabor Wavelet and Supervised Classification. IEEE Trans Medical imaging 2006, 25: 1214–1222. 10.1109/TMI.2006.879967View ArticleGoogle Scholar
- Elisa R, Renzo P: Retianl Blood Vessel Segmentation Using Line Operators and Support Vector Classification. IEEE Trans Medical imaging 2007, 26: 1357–1365. 10.1109/TMI.2007.898551View ArticleGoogle Scholar
- Ana MM, Aurelio C: Segmentation of Retinal Blood Vessels by Combining the Detection of Centerlines and Morphological Reconstruction. IEEE Trans Medical imaging 2006, 25: 1200–1213. 10.1109/TMI.2006.879955View ArticleGoogle Scholar
- Emmanuel JC, David LD: Continuous curvelet transform I. Resolution of the wavefront set. Appl Comput Harmon Anal 2005, 19: 162–197. 10.1016/j.acha.2005.02.003MATHMathSciNetView ArticleGoogle Scholar
- M'Elena MP, Alun DH, Simon AT, Anil AB, Kim HP: Segmentation of blood vessels from red-free and fluorescein retinal images. Med mage nal 2007, 11: 47–61.Google Scholar
- DRIVEdatabase [http://www.isi.uu.nl/Research/Databases/DRIVE]
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.