Segmentation of blood vessels from red-free and fluorescein retinal images
Introduction
The eye is a window to the retinal vascular system which is uniquely accessible for the non-invasive, in vivo study of a continuous vascular bed in humans. The detection and measurement of blood vessels can be used to quantify the severity of disease, as part of the process of automated diagnosis of disease or in the assessment of the effect of therapy. Retinal blood vessels have been shown to change in diameter, branching angles or tortuosity, as a result of a disease, such as hypertension (Stanton et al., 1995, King et al., 1996, Wong and McIntosh, 2005), diabetes mellitus (Skovborg et al., 1969) or retinopathy of prematurity (ROP) (Gelman et al., 2005). Furthermore, retinal arteriolar or venular changes predict development of hypertension (Wong and McIntosh, 2005, Ikram et al., 2006), new onset diabetes (Wong and McIntosh, 2005), progression to diabetic retinopathy (Klein et al., 2004) and development of diabetic renal disease (Wong et al., 2004). Thus a reliable method of vessel segmentation would be valuable for the early detection and characterisation of morphological changes.
Different techniques are used to acquire images of retinal blood vessels, two of them are of our interest in this work. A non-invasive technique is the retinal fundus photograph taken using a green filter. A variation of the latter, widely used clinically which is consider minimally invasive, implies the use of a topical mydriatic drop to dilate the pupil. These images are generally called red-free. A more invasive technique is fluorescein angiography which involves an intravenous injection of dye that increases the contrast of the blood vessels against the background. In the present work we will be referring only to monochrome images obtained by extracting the green band from RGB colour images. This band is chosen because it is known to show the improved visibility of the retinal blood vessels. Fig. 1 shows an example of two scanned negatives taken from the same eye before (red-free) and after the injection of fluorescein dye (fluorescein). There have been many studies on the detection of blood vessels in medical images in general but only some of them are related to retinal blood vessels in particular. Most of the work on segmentation of retinal images can be categorised into three approaches: those based on line or edge detectors with boundary tracing (Akita and Kuga, 1982, Wu et al., 1995), those based on matched filters, either 1-D profile matching with vessel tracking and local thresholding (Zhou et al., 1994, Gao et al., 1997, Gao et al., 2000, Tolias and Panas, 1998, Jiang and Mojon, 2003) or 2-D matched filters (Chaudhuri et al., 1989, Zana and Klein, 1997, Hoover et al., 2000), and those supervised methods which require manually labelled images for training (Sinthanayothin et al., 1999, Staal et al., 2004).
We have applied some of these methods but because of the large regional variations in intensity inherent in these retinal images and the very low contrast between vessels and the background, particularly in the red-free photographs, the results were disappointing. Techniques based on line or edge detectors lacked robustness in defining blood vessels without fragmentation and techniques based on matched filters were difficult to adapt to the variations of widths and orientation of blood vessels. Furthermore most of these segmentation methods are developed to work either on red-free or fluorescein images but not on both (Martinez-Perez, 2001).
In this paper we present a method based on multiscale analysis from which we obtain retinal blood vessel width approximation, size and orientation using gradient magnitude and maximum principal curvature of the Hessian tensor, two geometric features based upon the first and the second spatial derivatives of the intensity calculated for each different scale that give information about the topology of the image. We then use a multiple pass region growing procedure which progressively segments the blood vessels using the feature information together with spatial information about the eight-neighbouring pixels, obtaining in this way a segmented binary image. The algorithm works equally well with both red-free fundus images and fluorescein angiograms as will be shown in Sections 3 Results, 4 Validation of segmentation.
A first approximation to the segmentation of retinal blood vessel using this approach was previously presented (Martinez-Perez et al., 1999), where segmentation method was tested on a small image sample without any validation. An extension of this work is presented here and the method is now tested on two local databases and two public databases of complete manually labelled images (Hoover et al., 2000, Staal et al., 2004). We evaluate our segmentation using the public databases that have been also used by other authors for the same purpose (Jiang and Mojon, 2003, Staal et al., 2004). Validation of segmented vessel diameters and branching angles measurements are also made: between red-free against fluorescein images and between our algorithm and one of the public databases.
Section snippets
The segmentation method
Since retinal blood vessels have a range of different sizes it is convenient to introduce a measurement that varies within a certain range of scales. Multiscale techniques have been developed to provide a way to isolate information about objects in an image by looking for geometric features at different scales. Under this framework representing information at different scales is defined by convolving the original image I(x, y) with a Gaussian kernel G(x, y; s) of variance s2:
Results
A total of 114 images from four different sources have been tested. Two sets were provided by St. Mary’s Hospital, London. The first one corresponds to 20 images of 10 normotensive and 10 hypertensive subjects. The second one corresponds to 17 paired red-free and fluorescein images (a total of 34). The other two sources tested come from two public databases with 20 and 40 images, respectively (Hoover et al., 2000, Staal et al., 2004).
Both sets provided by St. Mary’s Hospital, London are retinal
Validation of segmentation
Before using these results in clinical studies, a validation of the accuracy of the segmentation is necessary. Since direct in vivo measurements of retinal blood vessels in humans are not feasible, three indirect validation studies were undertaken: (1) a comparison of our multiscale (MS) segmentation results with two public databases of hand-segmented images; (2) a comparison of automatic measurements of diameters taken from one MS segmented image and its corresponding hand-segmented image from
Conclusions
We have presented an algorithm to segment retinal blood vessels from both red-free and fluorescein images which combines: (1) the multiscale feature extraction that gives information about the width property of blood vessels that is independent of their orientation in the image, (2) two geometric properties of tube-like structures based on the first and second derivatives of intensity that give weights to pixels with a high probability of belonging to vessels, and (3) a multiple pass region
Acknowledgement
M. Elena Martinez-Perez would like to acknowledge to the Mexican National Council for Science and Technology (CONACYT) for financial support.
References (38)
- et al.
A computer method of understanding ocular fundus images
Pattern Recogn.
(1982) - et al.
Quantification and characterisation of arteries in retinal images
Comp. Meth. Program. Biom.
(2000) - et al.
Statistical methods for assessing agreement between two methods of clinical measurement
Lancet
(1986) - et al.
Microphotometry of the blood column and the light streak on retinal vessels in fundus photographs
Acta Ophthalmol.
(1986) The pyramid as a structure for efficient computation
- et al.
Detection of blood vessels in retinal images using two-dimensional matched filters
IEEE Trans. Med. Imag.
(1989) Ridges in Image and Data Analysis
(1996)- et al.
Multiscale vessel enhancement filtering
- Gao, X., Bharath, A., Hughes, A., Stanton, A., Chapman, N., Thom, S. 1997. Towards retinal vessel parametrisation. In:...
- et al.
Diagnosis of plus disease in retinopathy of prematurity using retinal image multiscale analysis
Invest. Ophth. Vis. Sci.
(2005)
Light paths in retinal vessel oximetry
IEEE Trans. Biomed. Eng.
Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response
IEEE Trans. Med. Imag.
Quantification of topological changes in retinal vascular architecture in essential and malignant hypertension
J. Hypertens.
Retinal vessel diameters and risk of hypertension: the rotterdam study
Hypertension
Adaptive local thresholding by verification-based multithreshold probing with application to vessel detection in retinal images
IEEE Trans. Pattern Recogn. Anal. Mach. Intell.
Arteriolar length–diameter (l:d) ratio: a geometric parameter of the retinal vasculature diagnostic of hypertension
J. Hum. Hypertens.
The relation of retinal vessel caliber to the incidence and progression of diabetic retinopathy – xix: the Wisconsin epidemiologic study of diabetic retinopathy
Arch. Ophthalmol.
The structure of images
Biol. Cybern.
Cited by (411)
SDDC-Net: A U-shaped deep spiking neural P convolutional network for retinal vessel segmentation
2023, Digital Signal Processing: A Review JournalExpACVO-Hybrid Deep learning: Exponential Anti Corona Virus Optimization enabled Hybrid Deep learning for tongue image segmentation towards diabetes mellitus detection
2023, Biomedical Signal Processing and ControlAn efficient registration-based approach for retinal blood vessel segmentation using generalized Pareto and fatigue pdf
2022, Medical Engineering and PhysicsHuman treelike tubular structure segmentation: A comprehensive review and future perspectives
2022, Computers in Biology and MedicineRecent trends and advances in fundus image analysis: A review
2022, Computers in Biology and Medicine