Locating Blood Vesls in Retinal Images
by Piece-wi Threshold Probing of a Matched Filter Respon Adam Hoover, Ph.D.+, Valentina Kouznetsova, Ph.D.+, Michael Goldbaum, M.D.∇+Electrical and Computer Engineering Department
∇Department of Ophthalmology
University of California, San Diego
La Jolla, CA 92093-0407
hoover or vkouznet@vision.ucsd.edu, mgoldbaum@ucsd.edu Key words: medical image processing, retinal imaging, blood vesl gmentation
ABSTRACT
We describe an automated method to locate and outline blood vesls in images of the ocular fundus. Such a tool should prove uful to eyecare specialists for purpos of patient screening, treatment evaluation, and clinical study. Our method differs from previously known methods in that it us local and global vesl features cooperatively to gment the vesl network. A comparison of our method a
gainst hand-labeled ground truth gmentations of five images yielded 65% nsitivity and 81% specificity. A previously known technique yielded 69% nsitivity and 63% specificity. For a baline, we also compared the ground truth against a cond hand labeling, yielding 80% nsitivity and 90% specificity. The numbers indicate our method improves upon the previously known technique, but that further improvement is still possible.
INTRODUCTION
Blood vesl appearance is an important indicator for many diagnos, including diabetes, hypertension, and arteriosclerosis. Vesls and arteries have many obrvable features, including diameter, color, tortuosity (relative curvature), and opacity (reflectivity). Artery-vein crossings and patterns of small vesls can also rve as diagnostic indicators. An accurate delineation of the boundaries of blood vesls makes preci measurements of the features possible. The measurements may then be applied to a variety of tasks, including diagnosis, treatment evaluation, and clinical study.
We describe an automated method to locate and outline blood vesls in images of the ocular fundus. With this tool, eyecare specialists can potentially screen larger populations for vesl abnor
malities. Preci measurements may be more easily recorded, for instance for evaluation of treatment or for clinical study. Obrvations bad upon such a tool would also be more systematically reproduceable.
Previous methods to gment blood vesls automatically have concentrated primarily on their local attributes. Vesls may be characterized by the expected color (reddish), shape (curvilinear), gradient (strength of boundary), and contrast (with background). Unfortunately, this description is not exclusive. For suitable ranges of the attributes, other image manifestations, such as the boundaries of the optic nerve and some hemorrhages and lesions, can exhibit the same local attributes as vesls.
Figure 1 shows an example retinal image, along with an image showing the result of the matched filter convolution described in [1]. The strength of the matched filter respon (MFR) is coded in greyscale: the darker a pixel, the stronger the respon. Notice that the strong respons in the center of the MFR image, which are obviously not vesl, are unfortunately much stronger than the respons on the left side of the MFR image, which are vesl. Therefore, applying a single global threshold does not provide adequate classification, as shown in Figure 2.
We propo a novel method to gment blood vesls that compliments local vesl attributes with region-bad attributes of the network structure. A piece of the blood vesl network is hypothesized by probing an area of the MFR image, iteratively decreasing the threshold. At each iteration, region-bad attributes of the piece are tested to consider probe continuation, and ultimately to decide if the piece is vesl. Pixels from probes that are not classified as vesl are recycled for further probing. The strength of this approach is that individual pixel labels are decided using local and region-bad properties.
RELATED WORK
meantodoPrevious methods to gment blood vesls generally fall into three categories: window-bad
[1,2,3], classifier-bad [4,5], and tracking-bad [6,7]. Window-bad methods, such as edge detection, estimate a match at each pixel for a given model against the pixel's surrounding window. In [1], the cross ction of a vesl in a retinal image was modeled by a Gaussian shaped curve, and then detected using rotated matched filters. In [2], a similar method was ud for artery detection in angiograms. In [3], a window surrounding a vesl was modeled by a neural network trained on ur-lected examples. The drawback of the methods is that the large-scale propertie
s of vesls (i.e., their network structure) must be ignored to insure computational feasibility.
Classifier-bad methods proceed in two steps. First, a low-level algorithm produces a gmentation of spatially-connected regions. The candidate regions are then classified as being vesl or not vesl. In [4], regions gmented by ur-assisted thresholding were classified as blood vesl or leakage according to their length to width ratio. In [5], regions gmented by the method in [1] were classified as vesl or not vesl according to many properties, including their respon to a classic operator designed to detect roads in aerial imagery [8]. The drawback of the methods is that the large-scale properties of vesls cannot be applied to the problem until after the low-level gmentation has already finished. Therefore, the properties cannot be ud to drive the gmentation, merely to evaluate it.
Tracking-bad methods utilize a profile model to step along and gment a vesl incrementally. In [6], a Hough transform is ud to locate the papilla in a retinal image. Vesl tracing proceeds iteratively from the papilla, halting when the respon to a one-dimensional (cross-ction) matched filter falls below a given threshold. In [7], a similar method was employed to detect vesls in coronary arteriograms,from ur-given starting points. One drawback to the approaches is their proclivity for termination at branch points, which are not well-modeled by one-dimensional filters. Ano
ther drawback is their reliance upon unsophisticated methods for locating starting points.
In [9], a method for tracking edge paths is ud to gment arteries in cineangiograms. Edge paths are modeled as Markov chains. A quential edge linking (SEL) algorithm is introduced to arch the possible t of paths for the best fit to the Markov model. The probabilities of the model are adjusted to reflect the properties of the desired path, such as the tolerance to local curvature. A strength of this approach is that the grouping operation works upon actual gradient values, as oppod to a thresholded respon. Therefore, a gmentation decision is not reached until an arbitrary number of pixels is available for classification. A drawback to the approach is that branches are not modeled, so that each branch must be traced and classified independently.
In this work, we propo a new method for gmenting blood vesls in a retinal image. The MFR image, computed as described in [1], is thresholded using a novel probing technique. The probe examines the image in pieces, testing a number of region-bad properties. If the probe decides a piece is vesl, then the constituent pixels are simultaneously gmented and classified. Contrasted against classifier-bad methods, our probing method allows a pixel to be tested in multiple region configurations before final classification. Contrasted against tracking-bad methods, our probing method is driven by a two-dimensional matched filter respon. Contrasted against [9], our probing
高中英语常用词组method is region-bad, and so naturally allows for multiple branches.
超低价qq业务自助下单平台
Figure 1:An example retinal image with obscured vesls (left) and matched filter respon (right). The respon is coded such that a darker value reprents a stronger respon.
ALGORITHM
The basic operation of the algorithm is to probe regions in a matched filter respon (MFR) image. During each probe, a t of criteria is tested to determine the threshold of the probe, and ultimately to decide if the area being probed (termed a piece) is blood vesl. A flowchart for the algorithm is shown in Figure 3. A queue of points is initialized, each of which will be ud for a probe. Upon a probe's completion, if the piece is determined to be vesl,then the endpoints of the piece are added to the queue.
In this way, different probes (and thus different
thresholds) can be applied throughout the image.
The following steps initialize a queue of pixels
that are to be ud as starting points for probing:•Convolve the matched filter described in [1] with the image, producing a matched filter respon
(MFR) image.
•Using a histogram of the MFR image, threshold the image such that > T THRESH pixels are above the threshold.
•Thin the thresholded image (for instance, using the algorithm given in [10], pg. 59).
kick•In the thinned image, era (relabel as background) all branchpoints, breaking up the entire foreground into gments that contain two endpoints each. Endpoints may be discovered as any pixel for which a traver of the eight bordering pixels in clockwi order yields only one foreground-to-background transition.
Similarly, branchpoints may be discovered as any
pixel for which the same traver yields more than
route66two transitions.
•Discard gments with less than T MIN pixels.
•All remaining endpoints are placed in the probe queue.
Each pixel in the probe queue is ud as a starting point for threshold probing. The probing is iterative; the iterations are ud to determine an appropriate threshold for the area being probed. The initial threshold is the MFR image value at the starting pixel. In each iteration, a region is grown from the start pixel, using a conditional paint-fill technique. The paint-fill spreads across all connecting pixels that are not
Figure 2:Matched filter respon thresholded at two different values. There is a strong overlap between true positive and fal positive respons.
Figure 3:Flowchart of algorithm.
already labeled and that are above the current threshold. Once the paint-fill is complete, the desired attributes of the grown region are tested. If the region pass the tests, then the threshold is decread by one, and a new iteration begins. Each probe iteration conducts the following tests:
•If the piece size (in pixels) exceeds T MAX, then the probe halts. This requires multiple pieces (and
thus potentially multiple thresholds) to gment the entire image. The effect is that the probe adapts to the local strength of the MFR image.•If the threshold reaches zero, then the probe halts.
This happens when probing a small area (even one pixel) interior to an area already classified as vesl.
•If the piece touches (on its border) more than one previously vesl-classified piece, then the probe halts. This is particularly uful for bridging gaps along vesls exhibiting weak MFR values.
•If the ratio border-pixels-touching-another-piece :
total-pixels-in-piece > T FRINGE, then the piece is fringing, and the probe halts. This prevents a probe from arching along the borders of vesl pieces already gmented.
•If the piece grows a loop, then the probe halts.
Loops are detected by thinning the piece, and counting the endpoints and branchpoints. If the number of endpoints exceeds the number of branchpoints by more than two, there is a loop.
This test prevents a probe from arching along circular MFRs, such as tho caud by some lesions and hemorrhages.
•If the ratio total-pixels-in-piece : branches-in-piece < T TREE, then the probe halts. This requires a piece to have a minimum span of vesl(s) per branch, and thus prevents over-branching down fal paths.
Once the probe is complete, if the resulting region has at least T MIN pixels, but less than T MAX pixels, then the region is labeled as vesl. The endpoints of the vesl piece are added to the queue. If the region is not determined to be vesl, then its pixels are left unlabeled. In either ca, the next point in the queue is lected for probing. When the queue is empty, the algorithm is complete.
EXPERIMENTS
Five retinal fundus slides were lected for testing the described method. Each slide was digitized to produce a 605 × 700 pixel image, 24-bits per pixel (standard RGB). All five images contain abnormalities that obscure or confu the blood vesl appearance. This lection was made for two reasons. First, most of the referenced methods have only been demonstrated upon normal vesl appearances, which are easier to discern. Second, some level of success with non-normal vesl appearances must be established to recommend clinical usage.
傻瓜的英文单词Each of the five images was carefully labeled by hand, to produce a ground truth gmentation of vesls. An example is shown in Figure 4. Each of the five images was procesd by the described algorithm, using the parameters T THRESH = 30800, T MIN =150, T MAX = 3500, T FRINGE = 0.3, and T TREE = 200. The values were lected after exploratory experiments, except for T THRESH, which was lected as the average number of pixels labeled as vesl in the ground truth images. An example result is shown in Figure 5. For comparison, each of the five images was globally thresholded using the same value of T THRESH. Figure 2 (right) shows this result for the example.
Each global-gmented result and probing-gmented result was compared against the ground truth, as follows. The percentage of pixels correctly gmented as vesl (true positive) was calculated as the number of pixels gmented as vesl that were within one pixel's distance of a pixel hand labeled as vesl, divided by the total number of pixels hand labeled as vesl. The percentage of pixels incorrectly gmented as vesl (fal positive) was calculated as the number of pixels gmented as vesl that were not within one pixel's distance of a pixel hand labeled as vesl, divided by the total number of pixels hand labeled as vesl. The tolerance of one pixel in distance was ud to help minimize measurement error. For all five images, the global-gmented re
sult had 69% nsitivity (true positive rate) and 63% specificity (37% fal positive rate). The probing-gmented result had 65% nsitivity and 81% specificity.
Figure 4: A hand-labeled gmentation of the
Figure 5:The result of threshold probing on the
CONCLUSIONS
The described method gments roughly two-thirds of the vesls in a retinal fundus image. Compared to a previously reported method [1], which us only a global threshold, the propod method produces roughly half the fal positive respons, and a slightly decread true positive respon. The latter is mainly attributable to the restriction upon the approach to produce connected vesl gmentations. A global threshold is likely to gment small groups of isolated pixels, as in Figure 2. Although such pixels may actually be correctly labeled, their utility for measurement is probably limited.
In order to explore our method of evaluation further, a cond person produced an additional t of hand-labeled ground truth for the five test images. This cond t of ground truth was compared to
the first t of ground truth exactly as described above, yielding 80% nsitivity and 90% specificity. This suggests some interesting conclusions. First, the vesls in the images lected for testing may in fact be too difficult to discern with 100% accuracy, so that our results must be viewed accordingly. In contrast, our method of evaluation may need to be changed so that competing hand-labeled ground truths score near perfect, thus more accurately reflecting the strength of automated approaches. Finally, we note that in either ca there is still measurable room for improvement.
ACKNOWLEDGMENTS
essay范文>konanThis work was supported by NIH Library of Medicine grant LM 05759-09.
REFERENCES
[1]S. Chaudhuri, S. Chatterjee, N. Katz, M. Nelson
连衣裙英语单词and M. Goldbaum, ``Detection of Blood Vesls in Retinal Images Using Two-Dimensional Matched Filters'', in IEEE Trans. on Medical Imaging, vol. 8, no. 3, Sep. 1989, pp. 263-269. [2]T. Pappas and J. Lim, ``A New Method for
Estimation of Coronary Artery Dimensions in Angiograms'', in IEEE Trans. On Acoustics, Speech, an
d Signal Processing, vol. 36, no. 9, Sep.
1988, pp. 1501-1513.
[3]R. Nekovei, Y. Sun, ``Back-Propagation Network
and its Configuration for Blood Vesl Detection in Angiograms'', in IEEE Trans. on Neural Networks, vol. 6, no. 1, Jan. 1995, pp. 64-72. [4]S. Tamura, K. Tanaka, S. Ohmori, K. Okazaki, A.
Okada and M. Hoshi, ``Semiautomatic Leakage Analyzing System for Time Series Fluorescein Ocular Fundus Angiography'', in Pattern Recognition, vol. 16, no. 2, 1983, pp. 149-162. [5] B. Cote, W. Hart, M. Goldbaum, P. Kube and M.
Nelson, ``Classification of Blood Vesls in Ocular Fundus Images'', technical report, Computer Science and Engineering Dept., University of California, San Diego, 1994.
[6]S. Tamura, Y. Okamoto and K. Yanashima,
``Zero-Crossing Interval Correction in Tracing Eye-Fundus Blood Vesls'', in Pattern Recognition, vol. 21, no. 3, 1988, pp. 227-233. [7]Y. Sun, ``Automated Identification of Vesl
Contours in Coronary Arteriograms by an Adaptive Tracking Algorithm'', in IEEE Trans. on Medical Imaging, vol. 8, no. 1, March 1989, pp.
78-88.
[8]M. Fischer, J. Tenenbaum and H. Wolf,
``Detection of roads and linear structures in low resolution aerial imagery using a multisource knowledge integration technique'', in Computer Graphics and Image Processing, vol. 15, no. 3, 1981, pp. 201-223.
[9]P. Eichel, E. Delp, K. Koral and A. Buda, ``A
phd是什么意思
Method for a Fully Automatic Definition of Coronary Arterial Edges from Cineangiograms'', in IEEE Trans. on Medical Imaging, vol. 7, no. 4, Dec. 1988, pp. 313-320.
[10]R. Jain, R. Kasturi and B. G. Schunck, Machine
Vision, McGraw-Hill, 1995.
