Products & Services

DSP implementation of a low-complexity algorithm for real-time automated vessel detection in images of the fundus of the human retina

DSP implementation of a low-complexity algorithm for real-time automated vessel detection in images of the fundus of the human retina
of 4
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
  DSP implementation of a low-complexity algorithmfor real-time automated vessel detection in imagesof the fundus of the human retina  Andrea Anzalone, Federico Bizzarri, Paolo Camera, Luca Petrillo, Marco Storace Biophysical and Electronic Engineering Dept., Univ. of GenoaVia Opera Pia 11a, I-16145, Genova, ItalyE-mail:  Abstract   —In this paper we propose an algorithm for real-time vessel segmentation in sequences of RGB images of thehuman retina. The algorithm is made up of two main blocksproviding vessel enhancement and thresholding, respectively.It is implemented on a DSP board for future embedding inophthalmology equipments. The obtained results (binary images)show a good trade-off between processing speed and accuracy,since the main structure of the vessel network is preserved andthe simplicity of the algorithm allows the DSP to process aboutten images per second. I. I NTRODUCTION The assessment of the characteristics of vessels of thehuman retina plays an important role in a variety of medicaldiagnoses [1], [2]. When the number of vessels in an imageis large, or when a high number of images is acquired,manual delineation of the vessels may become tedious or evenimpossible [2]. Moreover, during outpatient surgery on theretina, such as photoablation, the target is processing imagesequences to extract features and signal possible dangers oranomalies in real time. In this latter case, the availability of asystem embedded in ophthalmology equipments with a goodtrade-off between processing speed and accuracy would beprobably the best way to hit the target.One of the most signi Þ cant features in retina images isthe vessel caliber, that can be quite easily detected once thenetwork of veins, arteries and capillaries is extracted fromthe image. In other words, one of the main preliminary stepsfor feature extraction from the retina images is the vesseldetection. There are several categories of methods for vesselextraction [3] and two main groups of methods for vesselsegmentation in images of the retina. The  Þ rst group consistsof rule-based methods and comprises vessel tracking [4],matched  Þ lter responses [5], [6], grouping of edge pixels[7], model-based locally adaptive thresholding [8], topologyadaptive snakes [9], and morphology-based techniques [10],[13]. The second group concerns supervised methods, whichrequire manually labeled images for training [2].In this paper, we propose an algorithm suitable for DSPboards for the automated segmentation of vessels in two-dimensional RGB color images of the retina. These images,also known as fundus images, are acquired by taking pho-tographs of the posterior pole of the eye. The processing isperformed on  M   × N   pixels monochromatic red-free imagesobtained by extracting the green channel from the srcinalRGB images. As a matter of fact, the green channel of RGBfundus images gives the highest contrast between vessels andbackground [5].The algorithm is made up of two fundamental steps. The for-mer is devoted to vessel enhancement, while the latter providesa binary image by resorting to a proper thresholding procedure.The algorithm performances have been tested through the 40color fundus photographs of the database DRIVE [15]. Theextracted vessel structures have been compared with the vesseltrees manually extracted by people trained by an experiencedophthalmologist and available in the database. For the imple-mentation of the proposed algorithm we used the evaluationkit TS101 EZ-KIT Lite (300 MHz of clock frequency). Theobtained results show a good trade-off between processingspeed and accuracy. The main structure of the vessel networkis preserved and the simplicity of the algorithm allows theDSP board to process about ten  400 × 400  images per second.II. T HE ALGORITHM The proposed algorithm is made up of two fundamentalblocks. The  Þ rst block is devoted to the vessel enhancementoperation, while the second one provides a binary image byresorting to a proper thresholding procedure.  A. Block 1: vessel enhancement  To perform the vessel enhancement operation, we will adoptthe method  Þ rstly proposed [11], [12] and then employed[13] to process two-dimensional fundus images. The red-freeimage to be processed is viewed as a bi-dimensional function I  ( x,y )  de Þ ned over a rectangular domain  Ω ⊂ R 2 . The srcinof the axes  ( x,y )  is set in the center of   Ω . The function I  ( x,y )  represents a spatial distribution of grey levels, i.e. asurface of grey level intensities having a speci Þ c shape. Thevessel enhancement procedure, devoted to highlight geometrictubolar structures, can be achieved by evaluating the Hessianoperator  H  of the function  I  ( x,y )  properly convoluted with abi-dimensional Gaussian function with zero mean and standarddeviation  σ  (see Eq. (1)) G ( x,y ; σ ) = 12 πσ 2 exp  x 2 + y 2 2 σ 2   (1) 97 1-4244-0921-7/07 $25.00 © 2007 IEEE.  The convolution of the function  I   with the Gaussian  G is necessary as the evaluation of the second order spatialderivatives of   I   both increases the effects of the acquisitionnoise (that is certainly present in the image to be processed),and enhances high-frequency contents of the fundus imagethat are not useful for a good segmentation of vessels. Theconvolution operation is then nothing more than a low-pass Þ ltering operation. According to the Scale-Space Theory [14],the evaluation of   H ( I  ( x,y ) ⊗ G ( x,y ; σ ))  turns out to be H  =   L xx ( x,y ; σ )  L xy ( x,y ; σ ) L yx ( x,y ; σ )  L yy ( x,y ; σ )   (2) where  L xx ( x,y ; σ ) =  I  ( x,y ) ⊗ G xx ( x,y ; σ ) ,  L xy  =  L xy  = I  ⊗ G xy , and  L yy  =  I  ⊗ G yy .The eigenvalues of the Hessian  H  measure convexity andconcavity of the function  I  ( x,y ) ⊗ G ( x,y ; σ )  in the corre-sponding eigendirections. The expression of such eigenvaluesis reported in Eq. (3), where the dependence of the r.h.s. of such an expression on both the space variables and  σ  has beenomitted for the sake of conciseness. λ ± ( x,y ; σ ) = L xx  + L yy ±   ( L xx − L yy ) 2 + 4 L 2 xy 2  (3) At each point, the eigenvalue with the maximum absolutevalue is denoted as  Λ( x,y ; σ )  and the corresponding eigen-vector is orthogonal to the direction of maximum curvatureof the grey level. In the considered images, a high positivecurvature marks the presence of furrows in the surface  I  ( x,y ) ,i.e., vessels in the image, whereas negative curvature does notcorrespond to the presence of vessels [13]. Then, the processedimage can be obtained as follows: ˜ I  ( x,y ; σ ) =  max { 0 , Λ( x,y ; σ ) }  (4)The standard deviation  σ  is called  scale  and must beproperly set. The multi-scale algorithms take into account theprocessing results obtained with different scales. The resultsare usually accurate, but at the cost of high computationtimes. The alternative, adopted also in this paper, is to setthe scale heuristically, on the basis of the processing resultsof a suf  Þ ciently large set of images. Basically, the scale  Þ tsthe average vessel thickness in the considered images. At this point, the histogram of the grey levels of image ˜ I  ( x,y ; σ )  is stretched between 0 and 255.  B. Block 2: image binarization and cleaning In order to segment the vessels through image binarization,two strictly related parameters must be preliminarily set. The Þ rst one is a threshold grey level, say  Th , whereas the secondone is the fraction of image pixels with grey levels between0 and  Th . We choose heuristically the fraction at 0.9 andautomatically calculate  Th . So doing, the threshold dependson the image only, not on possible scalings on the imageluminosity level. Once the image  ˜ I   is binarized, it is useful todelete spurious elements not belonging to the vessel network.To this end, we adopted a simple algorithm, illustrated inFig. 1. We  Þ x a virtual grid made up of squares of   n × n (a) (b)(c) (d) Fig. 1. The two steps (by rows) of the cleaning algorithm. pixels and, for each square, we focus on the perimetric pixels.If such pixels are all black, we assume that the correspondingsquare contains either only background pixels or spuriouselements, not connected with the vessel structure. In bothcases, the whole square is set to black, thus removing thepossible spurious elements. This cleaning algorithm can beiterated by changing  n  or the virtual grid position, so as toaccurately clean the image, but at the cost of an increasingcomputational effort. We choose to iterate the algorithm onlytwice, with  n  = 10 . In the  Þ rst step, the grid completely coversthe image (see a detail in Fig. 1a) and some spurious elementsor not connected parts of vessels (see the grey squares inFig. 1a) are removed, as shown in Fig. 1b. In the second step,the grid is shifted by 5 pixel both horizontally and vertically(green grid in Fig. 1c) and other elements (see the grey squaresin Fig. 1c) are cleaned, as shown in Fig. 1d.III. DSP  IMPLEMENTATION The main adjustments made to the algorithm to  Þ t it to aDSP implementation are: 1) use of integer arithmetics, 2) useof intrinsic functions, 3) use of a unitary step function to avoidif-then-else control structures, 4) use of fast convolution.In particular, for the point 1), the srcinal-image grey levelsin binary format can be left-shifted of 10 or 7 bits in order touse 32 or 16 bits representations, respectively, without generateover ß ows. Concerning the point 2), besides the functions forregister handling, we used mainly one intrinsic function of the DSP board, that is the function  1 / √  x , both to calculatethe Hessian eigenvalues according to Eq. (3) and to performdivisions by numbers not multiples of 2.By exploiting the DMA channels and after code opti-mization, we managed to process about 10 images of size 400 × 400  per second. The use of a DSP board more suitedfor image processing applications should allow one to reducethe processing times and/or increase the image size. 98  (a) (b) (c) Fig. 2. Example of vessel extraction (see text). IV. E XPERIMENTAL RESULTS The processed fundus images are normal (non pathological)samples from the DRIVE (Digital Retinal Images for VesselExtraction) database [15].Figure 2 shows an example of image processing with inter-mediate results. These results have been obtained by applyingthe proposed algorithm to large-size images and without anyDSP-code optimization. Figure 2a is the green channel compo-nent of a DRIVE image, Fig. 2b is the intermediate image afterblock 1, Fig. 2c is the processed image. To have a qualitativemeasure of the accuracy of the results, we compared ourresults with the vessel network manually extracted from theimage (found in the database DRIVE) and calculated the falsepositive pixels (i.e., white pixels in the processed image thatdo not belong to manually extracted vessels) and the falsenegative pixels (i.e., black pixels in the processed image thatbelong to manually extracted vessels). Figure 3a shows themanually extracted vessels, whereas Fig. 3b shows the falsepositive (in green) and the false negative (in orange) pixels.Figure 3b witnesses the acceptable quality of the results, sincethe thinnest vessels (hardly visible also in the srcinal image)are not extracted, but the main vessel structure is preserved.We point out that the algorithm provides not only the vesseltree but also the edge of the Field Of Visual (FOV). Moreover,in the presence of pathologies (e.g., drusen) the algorithm would extract also other structures, besides the vessels, thatcould be partially or totally removed by the cleaning algorithm(or by higher-level procedures).Figure 4 shows further image processing results, concerningDRIVE images with different characteristics. These resultshave been obtained by applying the DSP-implemented versionof the algorithm to images of size  400  ×  400 . The greenchannel components of the original images are shown inthe left column and the  Þ nal vessel extraction is shown inthe right column. To have a quantitative evaluation of theaccuracy of the results, we resort to the Maximum Average Accuracy (MAA), a measure of performance based on twoimages: a reference binary image  ¯ I   - resulting from the manualsegmentation of a fundus image  I   performed by people trainedby an experienced ophthalmologist - and the binary image  ˆ I  - resulting from the proposed algorithm.The MAA evaluates the performance of the vessel detectionalgorithm in correspondence with the  N  FOV    pixels belonging(a)(b) Fig. 3. Quality of the vessel extraction (see text). 99  Fig. 4. Further examples of vessel extractions (by rows). to the FOV. This measure expresses the number of pixels thathave been correctly classi Þ ed with respect to  N  FOV    :MAA   = 1 −  j,k ∈ FOV  ¯ I  jk −  ˆ I  jk  N  FOV   ∈ [0 , 1]  (5)The MAA indices for the right column images of Fig. 4 are(from top to bottom): 0.915, 0.911, 0.921, 0.911. These resultsare acceptable [16], even if they could be further improved byde Þ ning proper supervised strategies to optimize the values of the parameters  σ  and  Th .V. C ONCLUDING REMARKS In this paper we proposed an algorithm for automatedvessel extraction from images of the human retina particularlysuited for DSP implementation. The image processing results will be provided to higher-level algorithms for the automateddetection of anomalies or dangers. The good trade-off betweenprocessing speed and accuracy should allow a quite easyembedding of a DSP circuit in ophthalmology equipments,in order to process sequences image sequences and signalpossible dangers or anomalies in real time during outpatientsurgery on the retina. A  CKNOWLEDGMENTS The authors would thank Dr. Federico Ricci (Departmentof Biopathology, University of Rome “Tor Vergata”) and Prof.Mauro Parodi for useful comments and discussions. This work was supported by MIUR, within the PRIN framework (projectno.  2004092944 003 ), and by University of Genoa.R EFERENCES[1] C.L. Tsai, C.V. Stewart, H.L. Tanenbaum, B. Roysam, “Model-BasedMethod for Improving the Accuracy and Repeatability of EstimatingVascular Bifurcations and Crossovers From Retinal Fundus Images,”  IEEE Trans. Inform. Technol. Biomed. , vol. 8, pp. 122–129, 2004.[2] J. Staal, M.D. Abramoff, M. Niemeijer, M.A. Viergever, B. van Gin-neken, “Ridge-Based Vessel Segmentation in Color Images of theRetina,”  IEEE Trans. Med. Imag. , vol. 23, pp. 501–509, 2004.[3] C. Kirbas, F. Quek, “A review of vessel extraction techniques andalgorithms,”  ACM Comp. Surv. , vol. 36, pp. 81-121, 2004.[4] A. Can, H. Shen, J.N. Turner, H.L. Tanenbaum, B. Roysam, “Rapidautomated tracing and feature extraction from retinal fundus imagesusing direct exploratory algorithms,”  IEEE Trans. Inform. Technol. Biomed. , vol. 3, pp. 125-138, 1999.[5] A. Hoover, V. Kouznetsova, M. Goldbaum, “Locating blood vessels inretina images by piecewise threshold probing of matched  Þ lter response,”  IEEE Trans. Med. Imag. , vol. 19, pp. 203–210, 2000.[6] L. Gang, O. Chutatape, S.M. Krishnan, “Detection and measurement of retinal vessels in fundus images using amplitude modi Þ ed second-orderGaussian  Þ lter,”  IEEE Trans. Biomed. Eng. , vol. 49, pp. 168-172, 2002.[7] A. Pinz, S. Berngger, P. Datlinger, A. Kruger, “Mapping the humanretina,”  IEEE Trans. Med. Imag. , vol. 17, pp. 606-619, Aug. 1998.[8] X. Jiang, D. Mojon, “Adaptive local thresholding by veri Þ cation- basedmultithreshold probing with application to vessel detection in retinalimages,”  IEEE Trans. PAMI  , vol. 25, pp. 131-137, Jan. 2003.[9] T. McInerney, D. Terzopoulos, “T-snakes: Topology adaptive snakes,”  Med. Imag. Anal. , vol. 4, pp. 73-91, 2000.[10] F. Zana, J.C. Klein, “Segmentation of vessel-like patterns using mathe-matical morphology and curvature evaluation,”  IEEE Trans. Image Proc. ,vol. 10, pp. 1010-1019, 2001.[11] Y. Sato  et al. , “3D multi-scale line  Þ lter for segmentation and visu-alization of curvilinear structures in medical images,”  Proc. CVRMed- MRCAS , vol. 1205, pp. 213–222, 1997.[12] C. Lorenz  et al. , “Multi-scale line segmentation with automatic estima-tion of width, contrast and tangencial direction in 2D and 3D medicalimages,”  Proc. CVRMed-MRCAS , vol. 1205, pp. 233–242, 1997.[13] M. Martinez-Perez  et al. , “Segmentation of retinal blood vessels basedon the second directional derivative and region growing,”  Proc. IEEE Int. Conf. Imag. , pp. 173–176, 1999.[14] T. Lindeberg,  Scale-Space Theory in Computer Vision , The Netherlands:Kluwer Academic Publishers, 1994.[15] [16] N. Niemeijer  et al. , “Comparative study of retinal vessel segmentationmethods on a new publicly available database,”  Proc. SPIE Medical  Imaging 2004: Image Processing , pp. 648–656, 2004. 100
Similar documents
View more...
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks