A system and method for robust segmentation of retinal vasculature
Patent Information
- Application Number
- IN202411075303
- Authority / Receiving Office
- IN · IN
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2024-10-04
- Publication Date
- 2026-07-14
- Estimated Expiration
- 2044-10-04
AI Technical Summary
Existing methods for automated retinal vessel segmentation struggle with effectively identifying thin vessels, handling complex vessels near the optic disc, and maintaining accuracy across varying imaging conditions.
A system and method that utilize a seven-dimensional feature vector to represent pixel characteristics in retinal fundus images, combined with a linear support vector machine for pixel classification, and enhanced image processing techniques such as adaptive histogram equalization and B-COSFIRE filters to improve segmentation accuracy.
The proposed solution achieves robust and accurate segmentation of retinal vasculature, efficiently detecting thin vessels and maintaining high accuracy across different imaging conditions, thereby improving diagnostic efficiency in ophthalmology.
Abstract
Description
FIELD OF THE INVENTIONThe present invention relates to the field of segmentation of the retinalvasculature. More particularly, the present invention relates to a system andmethod for robust segmentation of retinal vasculature.BACKGROUND OF THE INVENTIONRetinal imaging is a critical tool in modern ophthalmology, allowing for thedetailed examination of the vascular structures within the eye. Retinal bloodvessels are key indicators for diagnosing a wide range of eye-related diseases,such as diabetic retinopathy, glaucoma, and retinal artery occlusion. Inclinical practice, the accurate segmentation of these vessels is essential forearly detection, diagnosis, and treatment planning. However, due to theintricate and delicate nature of retinal vasculature, manual segmentation andanalysis are both time-consuming and prone to human error. Automatedsystems for retinal vessel segmentation have thus become a focal point inmedical imaging research, aiming to streamline the diagnostic process andreduce the burden on healthcare professionals.In the field of automated retinal vessel segmentation, several technologieshave emerged, each employing different methodologies to detect and mapretinal blood vessels from fundus images. Traditional methods often rely onstatistical texture analysis and probabilistic models to differentiate betweenblood vessels and the background. More recent advancements employ deeplearning and neural networks to enhance segmentation accuracy. However,despite these advancements, existing methods still struggle with certainlimitations, such as effectively identifying thin vessels, handling thecomplexity of vessels near the optic disc, and maintaining accuracy across arange of imaging conditions.US8787638B2 describes a method for diagnosing retinal disease usingstatistical texture analysis of retinal fundus images, employing techniquessuch as gray-level co-occurrence matrix (GLCM), run length matrix (RLM),and fractal and high-order spectra analysis. However, it lacks effectivesegmenting of the optic disc., also this invention faces challenges in thedetection of thin vessels from the images.WO2018237011A9 discloses a method for retinal blood vessel segmentationusing optical coherence tomography angiography (OCTA) images. It involvesthe preprocessing of images to calculate thresholds and improve contrast,followed by segmentation based on probability models, and refining theresults to label each pixel accurately. It also incorporates the techniques suchas combining multiple probability models and using supervised learningmodels for diagnosing conditions like diabetic retinopathy. However, it faceslimitations in effectively segmenting very fine or complex retinal bloodvessels, such as those near the optic disc. It relies on general probabilisticmodels and standard preprocessing techniques which may not sufficientlyaddress the challenges of accurately detecting and analyzing these delicatestructures which results on reduced accuracy and missed pathologicalfeatures.Therefore, there is a need for an improved system and method for robustsegmentation of retinal vasculature.OBJECT OF THE INVENTIONThe main object of the present invention is to provide a solution for robustsegmentation of the retinal vasculature.Another object of the present invention is to provide a solution that detectsthin blood vessels efficiently.Yet another object of the present invention is to provide a system and methodfor robust segmentation of retinal vasculature in which a seven-dimensionalfeature vector is constructed corresponding to every pixel in the retinal fundusimage for representing various characteristics of retinal blood vessels.Thereafter, a support vector machine is used to classify each pixel as a vesselor a background pixel.Still another object of the present invention is to provide a solution thatclassifies each pixel as a vessel or a non-vessel pixel to obtain the segmentedretinal vessels.SUMMARY OF THE INVENTIONThe present invention relates to a system and method for robust segmentationof retinal vasculature.In an embodiment, the present invention provides a system for robustsegmentation of retinal vasculature comprising a memory, a processorconnected to the memory. The processor is configured to receive, via a dataacquisition module, a retinal image from a user. The processor is furtherconfigured to enhance, via a filter unit, the retinal image to obtain a de-noisedretinal image. Further, adaptive histogram equalization is applied followed byvasculature enhancement by the B-COSFIRE filters. A centerline technique isalso applied on the green channel of the RGB retinal image. Then using acombination of seven features a linear support vector machine is used toclassify a pixel as a vessel or a background. The processor is furtherconfigured to display, via display unit, the plurality of segmented retinalvessels to the user.In another embodiment, the present invention provides a method for robustsegmentation of retinal vasculature comprises steps of, a) receiving, via a dataacquisition module of a processor, a retinal image from a user, b) enhancing,via a filter unit of the processor, the retinal image to obtain a de-noised retinalimage, c) applying, via the processor, adaptive histogram equalization and BCOSFIREfilters to obtain enhanced retinal vasculature, d) processing, via theprocessor; the green channel of the RGB retinal image to obtain thecenterlines in the retinal blood vessels, e) constructing, via the processor, aseven dimensional feature vector and using a linear support vector machinefor pixel classification, f) displaying, via a display unit, the plurality ofsegmented retinal vessels to the user.The above objects and advantages of the present invention will becomeapparent from the hereinafter set forth brief description of the drawings,detailed description of the invention, and claims appended herewith.BRIEF DESCRIPTION OF THE DRAWINGSAn understanding of the system and method for robust segmentation ofretinal vasculature of the present invention may be obtained by reference tothe following drawings:Figure 1 is a block diagram of the system for robust segmentation of retinalvasculature according to an embodiment of the present invention.Figure 2 is a tabular representation of the isotropic arrangement ofcoefficients of the system for robust segmentation of retinal vasculatureaccording to an embodiment of the present invention.Figure 3 is a flow diagram of the method for robust segmentation of retinalvasculature according to an embodiment of the present invention.Figure 4(a) is a pictorial view of an input image for the system and methodfor robust segmentation of retinal vasculature according to an embodiment ofthe present invention.Figure 4(b) is a pictorial view of an output image produced by the system andmethod for robust segmentation of retinal vasculature according to anembodiment of the present invention.Figure 5 is a tabular representation of comparison results of present inventionwith conventional methods according to an embodiment of the presentinvention.DETAILED DESCRIPTION OF THE INVENTIONThe present invention will now be described hereinafter with reference to theaccompanying drawings in which a preferred embodiment of the invention isshown. This invention may, however, be embodied in many different formsand should not be construed as being limited to the embodiment set forthherein. Rather, the embodiment is provided so that this disclosure will bethorough, and will fully convey the scope of the invention to those skilled inthe art.Many aspects of the invention can be better understood with references madeto the drawings below. The components in the drawings are not necessarilydrawn to scale. Instead, emphasis is placed upon clearly illustrating thecomponents of the present invention. Moreover, like reference numeralsdesignate corresponding parts through the several views in the drawings.Before explaining at least one embodiment of the invention, it is to beunderstood that the embodiments of the invention are not limited in theirapplication to the details of construction and to the arrangement of thecomponents set forth in the following description or illustrated in thedrawings. The embodiments of the invention are capable of being practicedand carried out in various ways. In addition, the phraseology and terminologyemployed herein are for the purpose of description and should not be regardedas limiting.The present invention now will be described hereinafter with reference to thedetailed description, in which some, but not all embodiments of the inventionare indicated. Indeed, the invention may be embodied in many different formsand should not be construed as limited to the embodiments set forth herein;rather, these embodiments are provided so that this disclosure will satisfyapplicable legal requirements. Like numbers refer to like elements throughout.The present invention is described fully herein with non-limiting embodimentsand exemplary experimentation.The present invention relates to a system and method for robust segmentationof retinal vasculature that de-noise the retinal image, enhances the de-noisedimage and utilizes support vector machine technique for obtaining thesegmented retinal vessels.In an embodiment, the present invention provides a system (100) for robustsegmentation of retinal vasculature comprising a data storage unit (101), aprocessor (102) connected to the data storage unit (101). The processor (102)is configured to receive, via a data acquisition module (103), a retinal imagefrom a user. The processor (102) is further configured to enhance, via a filterunit (104), the retinal image to obtain a de-noised retinal image. Theprocessor (102) is further configured to apply a set of techniques on the de-noised retinal image for identifying a plurality of thin retinal vessels in the denoisedretinal image that yields a plurality of centerlines. The processor (102)is further configured to construct a seven dimensional feature vector and usesa linear support vector machine of pixel classification. The processor (102) isfurther configured to display, via display unit (105), the plurality of segmentedretinal vessels to the user.In another embodiment, the present invention provides a method (300) forrobust segmentation of retinal vasculature comprises steps of, a) receiving, viaa data acquisition module (103) of a processor (102), a retinal image from auser, b) enhancing, via a filter unit (104) of the processor (102), the retinalimage to obtain a de-noised retinal image, c) applying, via the processor (102),a set of techniques on the de-noised retinal image for identifying a plurality ofthin retinal vessels in the de-noised retinal image that yields a plurality ofcenterlines; d) processing, via the processor (102), a seven dimensional featurevector using a linear support vector machine for pixel classification; and e)displaying, via a display unit (105), the plurality of segmented retinal vesselsto the user.Referring to Figure 1, a block diagram of the system (100) for robustsegmentation of retinal vasculature, according to embodiment of the presentinvention. The system (100) comprises a data storage unit (101) and aprocessor (102) connected to the data storage unit (101). The data storage unit(101) comprises one or more instructions for implementing the features of thepresent invention. The processor (102) is configured to receive, via a dataacquisition module (103), a retinal image from a user. The data acquisitionmodule (103) obtains the retinal image from a database. The retinal image is ared, green, blue (RGB) retinal image.The processor (102) is further configured to enhance, via a filter unit (104), theretinal image to obtain a de-noised retinal image.The filter unit (104) of the processor (102) is configured to produce the de-noised retinal image by de-noising a green channel of the RGB retinal imagefrom a left or forward weighted fractional integral operator and enhance acontrast level via an adaptive histogram equalization technique. The filter unit(104) includes a fractional filter, adaptive histogram equalization, barcombination of shifted filter responses (B-COSFIRE) filter.The left / forward weighted fractional integral of order a > 0 of a functionf(t) with respect to another function z(t) and weight w(t) is defined asfollows in Equation (1): Equation (1)Substituting z(x) = px + q; p, q ϵ R - {0}, w(t) = eat , a > 0 and xl = 0 intothe above equation and discretizing it (refer to Equation (2)),Equation (2)For n = 2, the coefficients ck are given by, c0 =EquationUsing these coefficients, an 5 x 5 isotropic filter is produced asin Table 1 depicted in Figure 2.The contrast of the image denoised using the 5 x 5 isotropic filter is producedas in Table 1 depicted in Figure 2, is then enhanced using adaptive histogramequalisation. This results in the pre-processed image to be used for vasculatureenhancement by B-COSFIRE filters.The process of applying the B-COSFIRE filter involves a series of steps,starting with convolving the preprocessed image with the difference ofGaussian (DoG) filter, shifting the weighted DoG or blurred responses, andapplying point-wise weighted geometric mean. In this context, the DoGfunction, denoted as DoGσ (x, y), is defined as follows (refer to Equation 3): Equation (3)Here, σ stands for the standard deviation of the Gaussian function. A value of0.5 is chosen as the ratio of the standard deviations for the two Gaussianfunctions. The variables (x, y) refer to the position of a pixel in the image. Fora grayscale image I, the response of the DoG filter, denoted as cσ(x, y), iscomputed by convolving the DoG kernel, DoGσ with the image I (refer toEquation (4)). Equation (4)The notation |.| + signifies the process of half-wave rectification, whereinnegative values are set to zero. cσ(x, y), highlights all the vascular and nonvascularstructures in a retinal image. The next step is to retain only thevascular structures and eliminate other non-vascular structures, such as theoptic disc, lesion, and macula. The key idea used in the B-COSFIRE filter isthat out of all the linear neighborhoods of a vessel pixel, the neighborhoodalong the vessel has the lowest average intensity. As a result, for every pixel,an average intensity around its linear neighborhood is found in multipleorientations, and the lowest average is retained as the response of the BCOSFIREfilter for that pixel. The average intensity around a specific linearneighborhood of a pixel is the geometric mean of the blurred responses of thepixels in that linear neighborhood. For a pixel (x, y), let Sθ = {(ρi , φi ) | i =1, … , n} denote the polar coordinates of the pixels in its predefined linerneighborhood. Here, θ is the angle with respect to the horizontal axis used todefine the orientation of the linear neighborhood of the pixel underconsideration. Then, the average intensity in that linear neighborhood iscomputed as: (refer to Equation (5))Equation (5)Here, spi,φi (x, y) is the blurred response of the ith neighbor of (x, y) in itslinear neighborhood Sθ , and is defined as follows (refer to Equation (6))spi,φi (x, y) = maxx'y' {cσi-Equation (6)Where-3σ' ≤ x', y' ≤ 3σ',σ' = σ0' + aρi ,Δxi = -ρi cos φi,Δyi = -ρi sin φi ,Gσ' is the Gaussian function with standard deviation σ', σ0' and a areconstants. Here, | . |t represents the response after applying thresholding at thelevel of t, where t ranges from 0 to 1. Likewise, using the B-COSFIRE filters,responses are obtained in multiple orientations (θ ϵ {0°,15°,30°, ... 165°}),and the minimum response is retained as the final filter response at that pixel.This is used in B-COSFIRE filters to detect linear patterns. Nevertheless, thereare two variations of the B-COSFIRE filters, namely, symmetric B-COSFIREand asymmetric B-COSFIRE filters, the present invention utilizes bothvariations of the B-COSFIRE filters to enhance the retinal vasculature.The processor (102) is further configured to apply a set of techniques on thegray scale (green channel) retinal image for identifying a plurality ofcenterlines. This is done by locating the pixels that are local minima acrossthe retinal vasculature and are locally connected. Alongside, inimplementation, the processor (102) is configured to construct a sevendimensionalfeature vector from the RGB retinal image for representing aplurality of characteristics of the retinal vessels. The processor (102) is furtherconfigured to classify each pixel in the retinal image, via a linear supportvector machine classifier, using the seven-dimensional feature vector as avessel or a non-vessel pixel. The first three components of the sevendimensional feature vector are the red, green, and the blue channel of theRGB retinal image. The next two components are the response of thesymmetric and the asymmetric B-COSFIRE filters. The last two componentsare related to the length and the average intensity of the centerline pixels.The processor (102) is further configured to display, via display unit (105), theplurality of segmented retinal vessels to the user for the method applied.Referring to Figure 3, a flow diagram of the method (300) for robustsegmentation of retinal vasculature, according to an embodiment of thepresent invention.The method (300) for robust segmentation of retinal vasculature comprises aplurality of steps, and in an implementation, the method (300) is implementedon the system (100).At step (a), the method (300) comprises receiving, via a data acquisitionmodule (103) of a processor (102), a retinal image from a user. The dataacquisition module (103) obtains the retinal image from a database. Theretinal image is a (red, green, blue) RGB retinal fundus image.At step (b), the method (300) comprises enhancing, via a filter unit (104) ofthe processor (102), the retinal image to obtain a de-noised retinal image. Thede-noised retinal image is produced by de-noising a green channel of the RGBretinal image from a left or forward-weighted fractional integral operator.At step (c), the method (300) comprises applying, via the processor (102),adaptive histogram equalization and B-COSFIRE filters to obtain enhancedretinal vasculature.At step (d), method (300) comprises processing, via the processor (102), thegreen channel of the RGB retinal image to obtain the centerlines in the retinalblood vessels.At step (e), method (300) comprises constructing, via the processor, a seven-dimensionalfeature vector and using a linear support vector machine for pixelclassification.At step (f), the method (300) comprises displaying, via a display unit (105),the plurality of segmented retinal vessels to the user.EXAMPLE 1Experimentation AnalysisThe present invention provides the system (100) and method (300) for robustsegmentation of retinal vasculature, which takes the RGB retinal image asdepicted in Figure 4(a) as an input and returns segmented vessels as an outputas depicted in Figure 4(b). In the output image, the retinal vessels arehighlighted as white pixels, and the rest of the background is black.Further, the comparison results of present invention with conventionalmethods [R1, R2, and R3] is depicted in Table 2 of the Figure 5. Also, thepresent invention used other metrics like sensitivity (Se), specificity (Sp), localaccuracy (Loc-Aacc), Mathews correlation coefficient (MCC), geometricmean (GM) and F1 score (F1). All metrics are evaluated on a percentagescale.[R1]: N. Strisciuglio, G. Azzopardi, M. Vento, and N. Petkov, "Supervisedvessel delineation in retinal images with automatic selection of b-cosfire filters," Machine Vision and Applications, vol. 27, pp. 1137-1149, 2016.[R2]: S. Thangaraj, V. Periyasamy, and R. Balaji, "Retinal vesselsegmentation using neural network," IET Image Processing, vol. 12, no. 5,pp. 669 678, 2018.[R3]: B. S. Tchinda, D. Tchiotsop, M. Noubom, V. Louis-Dorr, and D. Wolf,"Retinal vessel segmentation using edge detection filters and neural network,"Informatics in Medicine Unlocked, vol. 23, p. 100521, 2021.Therefore, the present invention provides a system (100) and method (300) forrobust segmentation of the retinal vasculature, that detects thin blood vesselsefficiently and in which a seven-dimensional feature vector is constructed forrepresenting various characteristics of retinal blood vessels and further classifyeach pixel as a vessel or a non-vessel pixel to obtain the segmented retinalvessels.Many modifications and other embodiments of the invention set forth hereinwill readily occur to one skilled in the art to which the invention pertains,having the benefit of the teachings presented in the foregoing descriptions andthe associated drawings. Therefore, it is to be understood that the invention isnot to be limited to the specific embodiments disclosed and that modificationsand other embodiments are intended to be included within the scope of theappended claims. Although specific terms are employed herein, they are usedin a generic and descriptive sense only and not for purposes of limitation.The foregoing description of embodiments of the invention has beenpresented for purposes of illustration and description. It is not intended to beexhaustive or to limit the invention to the precise form disclosed, andmodifications and variations are possible in light of the above teachings ormay be acquired from practice of the invention. The embodiments werechosen and described in order to explain the principals of the invention and itspractical application to enable one skilled in the art to utilize the invention invarious embodiments and with various modifications as are suited to theparticular use contemplated.
Claims
1. A system (100) for robust segmentation of retinal vasculature comprising: a data storage unit (101); a processor (102) connected to the data storage unit (101); wherein: the processor (102) is configured to: i) receive, via a data acquisition module (103), a retinal image from a user; ii) enhance, via a filter unit (104), the retinal image to obtain a denoised retinal image; iii) apply a set of techniques on the de-noised retinal image for enhancing the retinal vasculature; iv) process a green channel of the red, green, blue (RGB) retinal image to yield a plurality of centerlines; v) construct a seven-dimensional feature vector using the red, green, and blue channels of the RGB retinal image, along with the symmetric and asymmetric response of the set of techniques, along with the average intensity and length of the centerline segments and the seven-dimensional feature vector trains a support vector machine to classify each pixel as a vessel or non-vessel pixel; vi) obtain a plurality of segmented retinal vessels upon the classification of each pixel; and vii) display, via a display unit (105), the plurality of segmented retinal vessels to the user.
2. The system (100) for robust segmentation of retinal vasculature, as claimed in claim 1, wherein said data acquisition module (103) obtains the retinal image from a database.
3. The system (100) for robust segmentation of retinal vasculature as claimed in claim 1, wherein said retinal image is a red, green, blue (RGB) retinal image.
4. The system (100) for robust segmentation of retinal vasculature as claimed in claim 3, wherein said the filter unit (104) of the processor (102) is configured to produce the de-noised retinal image by denoising a green channel of the RGB retinal image from a left or forward weighted fractional integral operator and enhance a contrast level via an adaptive histogram equalization technique.
5. The system (100) for robust segmentation of retinal vasculature as claimed in claim 1, wherein said filter unit (104) include a fractional filter, bar combination of shifted filter responses (B-COSFIRE) filter.
6. The system (100) for robust segmentation of retinal vasculature as claimed in claim 1, wherein said set of techniques include a support vector machine and an artificial intelligence technique, adaptive histogram equalization and B-COSFIRE filters.
7. The system (100) for robust segmentation of retinal vasculature as claimed in claim 1, wherein upon obtaining the de-noised retinal image, said processor (102) is configured to: a) construct a seven-dimensional feature vector from the RGB retinal image for representing a plurality of characteristics of the plurality of retinal vessels; and b) classify, via a linear support vector machine classifier, each pixel from a plurality of pixels from the seven-dimensional feature vector as a vessel or a non-vessel pixel.
8. The system (100) for robust segmentation of retinal vasculature as claimed in claim 1, wherein said processor (102) detects the plurality of thin retinal vessels via a centerline detection technique.
9. A method (300) for robust segmentation of retinal vasculature comprises steps of: a) receiving, via a data acquisition module (103) of a processor (102), a retinal image from a user; b) enhancing, via a filter unit (104) of the processor (102), the retinal image obtained from step (a), to obtain a de-noised retinal image; c) applying, via the processor (102), a set of techniques on the denoised image, to obtain enhanced retinal vasculature; d) processing, via the processor (102), a green channel of the red, green, blue (RGB) retinal image from the enhanced retinal vasculature to obtain the centerlines in the retinal blood vessels; e) constructing, via the processor, a seven-dimensional feature vector and using a linear support vector machine for pixel classification that results in plurality of segmented retinal vessels; and f) displaying, via a display unit (105), the plurality of segmented retinal vessels to the user.
10. The method (300) for robust segmentation of retinal vasculature as claimed in claim 9, wherein said data acquisition module (103) obtains the retinal image from a database.
11. The method (300) for robust segmentation of retinal vasculature as claimed in claim 9, wherein said retinal image is a red, green, blue (RGB) retinal image.
12. The method (300) for robust segmentation of retinal vasculature as claimed in claim 11, wherein said de-noised retinal image is produced by de-noising a green channel of the RGB retinal image from a left or forward weighted fractional integral operator and enhancing a contrast level via an adaptive histogram equalization technique.
13. The method (300) for robust segmentation of retinal vasculature as claimed in claim 9, wherein said set of techniques includes a support vector machine technique, adaptive histogram equalization and BCOSFIRE filters.
14. The method (300) for robust segmentation of retinal vasculature as claimed in claim 9, wherein upon obtaining the de-noised retinal image, the method (300) comprises steps of: a) constructing, via a processor (102), a seven-dimensional feature vector from the de-noised retinal image for representing a plurality of characteristics of the plurality of retinal vessels; and b) classifying, via a linear support vector machine classifier of the processor (102), each pixel from a plurality of pixels from the seven-dimensional feature vector as a vessel or a non-vessel pixel.
15. The method (300) for robust segmentation of retinal vasculature as claimed in claim 9, wherein said plurality of thin retinal vessels is detected by via a centerline detection technique which is further processed to construct two features in the seven dimensional feature vector.
