Automatic Optic Disc Detection From Retinal ... - Semantic Scholar

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 00, NO. 00, 2010

Automatic Optic Disc Detection From Retinal Images by a Line Operator

1

2

Shijian Lu*, Member, IEEE, and Joo Hwee Lim, Member, IEEE

3

21 22

Abstract—Under the framework of computer-aided eye disease diagnosis, this paper presents an automatic optic disc (OD) detection technique. The proposed technique makes use of the unique circular brightness structure associated with the OD, i.e., the OD usually has a circular shape and is brighter than the surrounding pixels whose intensity becomes darker gradually with their distances from the OD center. A line operator is designed to capture such circular brightness structure, which evaluates the image brightness variation along multiple line segments of specific orientations that pass through each retinal image pixel. The orientation of the line segment with the minimum/maximum variation has specific pattern that can be used to locate the OD accurately. The proposed technique has been tested over four public datasets that include 130, 89, 40, and 81 images of healthy and pathological retinas, respectively. Experiments show that the designed line operator is tolerant to different types of retinal lesion and imaging artifacts, and an average OD detection accuracy of 97.4% is obtained. Index Terms—Computer-aided diagnosis, line operators, optic disc (OD) detection, retinal image analysis.

23

I. INTRODUCTION

24

UTOMATIC optic disc (OD) detection from retinal images is a very important task in ocular image analysis [1], [2] and computer-aided diagnosis of various types of eye diseases [3]–[5]. It is often a key step for the detection of other anatomical retinal structures, such as retinal blood vessels and macula [1], [6], [7], [8]. More importantly, it helps to establish a retinal coordinate system that can be used to determine the position of other retinal abnormalities, such as exudates, drusen, and hemorrhages [9], [10]. Some OD detection techniques have been reported in the literature. The early techniques make use of different types of ODspecific image characteristics. In particular, some techniques search for the brightest regions [11], [12] or regions with the highest image variation [13], [14] resulting from the bright OD and the dark blood vessels within the OD. The limitation of these methods is that many retinal images suffer from various types of

25 26 27 28 29 30 31 32 33 34 35 36 37 38 39

A

retinal lesion, such as drusen, exudates, microaneurysms, and hemorrhage, and imaging artifacts, such as haze, lashes, and uneven illumination (as illustrated in Figs. 9–11) that often produce brighter regions or regions with higher image variation compared with the OD. Several OD detection techniques make use of anatomical structures among the OD, macula, and retinal blood vessels. For example, some methods are based on the anatomical structure that all major retinal blood vessels radiate from the OD [15]–[18]. Some other methods make use of the relative position between the OD and the macula that often varies within a small range [19], [20]. Compared with the image characteristics, the anatomical structures are more reliable under the presence of retinal lesion and imaging artifacts. However, the extraction of either retinal blood vessels or the macula is often a nontrivial task by itself. This paper presents a line operator that is designed to locate the OD from retinal images accurately. Line operators have been used to locate linear structures from different types of images. For example, Zwiggelaar et al. used a line operator to detect linear structures from mammographic images [21], where a line strength is evaluated by the difference between the largest average image intensity along one oriented line segment and the average image intensity within a local neighborhood window. Ricci and Perfetti [22] used a similar line operator to detect the linear structures that are associated with the retinal blood vessels. Our proposed line operator is designed to capture the circular brightness structure associated with the OD. In particular, it evaluates the image variation along multiple oriented line segments and locates the OD based on the orientation of the line segment with the maximum/minimum variation. Fig. 1(a) shows an example of retinal image in DRIVE project’s dataset [16], and Fig. 1(b) shows an image that simulates the circular brightness structure associated with the OD. As shown in Fig. 1, the OD has a specific brightness variation pattern where the image variation along Lc in Fig. 1(b) across the OD center usually reaches the maximum, whereas that along Lt orthogonal to Lc reaches the minimum. The proposed method has several advantages. First, the designed line operator is tolerant to the retinal lesion and various types of imaging artifacts that most image-characteristics-based methods cannot handle properly. The tolerance to the imaging artifacts and retinal lesion can be explained by the proposed line operator that is designed to capture the unique circular brightness structure associated with the OD. Second, the designed line operator is stable and easy for implementation. It requires neither the retinal blood vessel nor the macula information. Third, the designed line operator can be extended for macula detection

IE E Pr E oo f

4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

1

Manuscript received March 16, 2010; revised July 22, 2010; accepted September 22, 2010. Date of publication; date of current version. Asterisk indicates corresponding author. ∗ S. Lu is with the Department of Computer Vision and Image Understanding, Institute for Infocomm Research, Agency for Science, Technology and Research (A*STAR), Singapore 138632 (e-mail: [email protected]). J. H. Lim is with the Department of Computer Vision and Image Understanding, Institute for Infocomm Research, Agency for Science, Technology and Research (A*STAR), Singapore 138632 (e-mail: [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TBME.2010.2086455

0018-9294/$26.00 © 2010 IEEE

40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88

2


Fig. 1. Circular brightness structure associated with the OD. (a) Example of retinal image in DRIVE project’s dataset with OD labeled by a bold black circle. (b) Simulated circular brightness structure: L c crossing the OD center and L t orthogonal to L c are added to illustrate the line segments along which the image variation reaches the maximum and the minimum.

Q1

Fig. 3. Example line operator that uses 20 oriented line segments and set the line length p at 21.

IE E Pr E oo f

smoothing filter [24] that combines geometric closeness and photometric similarity as follows: ∞ ∞ −1 h(x) = k (x) f (ξ)c(ξ, x)s(f (ξ); f (x))dξ (1) −∞

−∞

with the normalization factor ∞ ∞ k(x) = c(ξ, x)s(f (ξ); f (x))dξ −∞

Fig. 2. Retinal image preprocessing. (a) Lightness of the example retinal image in LAB color space. (b) Enhanced retinal image by bilateral smoothing where multiple crosses along a circle label the pixels to be used to illustrate the image variation along multiple oriented line segments.

89 90 91 92 93 94

with little adaptation. Experiments over four public datasets verify its superior performance. The rest of this paper is organized as follows. Section II describes the proposed OD detection technique. Experimental results are then described and discussed in Section III. Some concluding remarks are finally drawn in Section IV.

95

II. PROPOSED METHOD

96

This section presents the proposed OD detection technique. In particular, we divide this section into four subsections, which deal with the retinal image preprocessing, designed line operator, OD detection, and discussion, respectively.

97 98 99

100

A. Retinal Image Preprocessing

101

Retinal images need to be preprocessed before the OD detection. As the proposed technique makes use of the circular brightness structure of the OD, the lightness component of a retinal image is first extracted. We use the lightness component within the LAB color space, where the OD detection usually performs the best [23]. For the retinal image in Fig. 1(a), Fig. 2(a) shows the corresponding lightness image. The retinal image is then smoothed to enhance the circular brightness structure associated with the OD. We use a bilateral

102 103 104 105 106 107 108 109

110 111

112

(2)

−∞

where f (x) denotes the retinal image under study. c(ξ; x) and s(f (ξ), f (x)) measure the geometric closeness and the photometric similarity between the neighborhood center x and a nearby point ξ. We set both c(ξ; x) and s(f (ξ), f (x)) by Gaussian functions. The geometric spread σd and the photometric spread σr of the two Gaussian functions are typically set at 10 and 1 as reported in [24]. For the retinal image in Fig. 2(a), Fig. 2(b) shows the filtered retinal image.

113 114 115 116 117 Q2 118 119 120

B. Designed Line Operator

121

A line operator is designed to the detect circular regions that have similar brightness structure as the OD. For each image pixel at (x, y), the line operator first determines n line segments Li , i = 1, . . . , n of specific length p (i.e., number of pixels) at multiple specific orientations that center at (x, y). The image intensity along all oriented line segments can thus be denoted by a matrix I(x, y)n ×p , where each matrix row stores the intensity of p image pixels along one specific line segment. Fig. 3 shows an example of line operator that uses 20 oriented line segments and sets the line length p = 21. As shown in Fig. 3, each line segment Li at one specific orientation can be divided into two line segments Li,1 and Li,2 of the same length (p − 1)/2 by the image pixel under study (i.e., the black cell in Fig. 3). The image variation along the oriented line segments can be estimated as follows:

122

Di (x, y) = fm dn (IL i , 1 (x, y)) − fm dn (IL i , 2 (x, y)) , i = 1, . . . , n

(3)

123 124 125 126 127 128 129 130 131 132 133 134 135 136

LU AND LIM: AUTOMATIC OPTIC DISC DETECTION FROM RETINAL IMAGES BY A LINE OPERATOR

Fig. 4. Image variation along multiple oriented line segments: Each graph shows the image variation vector D(x, y) of one retinal image pixel labeled by a cross along the circle in Fig. 2(b).

138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163

where fm dn (·) denotes a median function. fm dn (IL i , 1 (x, y)) and fm dn (IL i , 2 (x, y)) return the median image intensity along Li,1 and Li,2 , respectively. D = [D1 (x, y), . . . , Di (x, y), . . . , Dn (x, y)] is, therefore, a vector of dimension n that stores the image variations along n-oriented line segments. The orientation of the line segment with the maximum/minimum variation has specific pattern that can be used to locate the OD accurately. For retinal image pixels, which are far away from the OD, the orientation of the line segment with the maximum/minimum variation is usually arbitrary, but for those around the OD, the image variation along Lc [labeled in Fig. 1(b)] usually reach the maximum, whereas that along Lt reaches the minimum. Fig. 4 shows the image variation vectors D(x, y) of eight pixels that are labeled by crosses along a circle shown in Fig. 2(b). Suppose that there is a Cartesian coordinate system centered at the OD, as shown in Fig. 2(b). For the retinal image pixels in quadrants I and III, the image variations along the 1st–10th [i.e., Lt in Fig. 1(b)] and the 11th–20th (i.e., Lc ) line segments labeled in Fig. 3 reach the minimum and the maximum, respectively, as shown in Fig. 4. But for the retinal image pixels in quadrants II and IV, the image variations along the 1st–10th and the 11th–20th line segments instead reach the maximum and the minimum, respectively. An orientation map can, therefore, be constructed based on the orientation of the line segment with the maximum (or minimum) variation as follows: O(x, y) = argmax D(x, y)

(4)

i

164 165 166 167

Fig. 5. Orientation map of the retinal image in Fig. 2(b). (a) Gray orientation map that is determined by using (4). (b) Binary orientation map that is determined by using (5).

where n refers to the number of the oriented line segments used in the line operator. For the retinal image in Fig. 1(a), Fig. 5(a) and (b) shows the determined gray orientation map and binary orientation map, respectively. As shown in Fig. 5(a), for retinal image pixels in quadrants I and III around the OD, the orientation map is a bit dark because the orientation of the line segment with the maximum variation usually lies between 1 and (n/2) + 1. However, for retinal image pixels in quadrants II and IV, the orientation map is bright because the orientation of the line segment with the maximum variation usually lies between n/2 and n. The binary orientation map in Fig. 5(b) further verifies such orientation pattern. The OD will then be located by using the orientation map to be described in the following.

IE E Pr E oo f

137

3

where D(x, y) denotes the image variation vector evaluated in (3). In addition, a binary orientation map can also be constructed by classifying the orientation of the line segment with the maximum variation into two categories as follows: Q(x, y) =

−1, 1,

if argmax D(x, y) < i

otherwise

n +1 2

168 169 170 171 172 173 174 175 176 177 178 179 180 181

C. OD Detection

182

We use a line operator of 20 oriented line segments because line operators with more line segments have little effect on the orientation map. The line length p is set as follows:

183

p = kR

(6)

where R denote the radius of the central circular region of retinal images as illustrated in Fig. 1(a). Parameter k controls the line length, which usually lies between 1/10 and 1/5 based on the relative OD size within retinal images [25]. The use of R incorporates possible variations of the image resolution. The specific pattern within the orientation map is captured by a 2-D circular convolution mask shown at the upper left corner of two peak images in Fig. 6. As shown in Fig. 6, the convolution mask can be divided into four quadrants, where the cells within quadrants I and III are set at −1, whereas those within quadrants II and IV are set at 1 based on the specific pattern within the orientation map. An orientation map can thus be converted into a peak image as follows: P (x, y) =

x 0 +m

y 0 +m

184 185

M (x, y)O(x, y)

186 187 188 189 190 191 192 193 194 195 196 197 198

(7)

x=x 0 −m y =y 0 −m

(5)

where (x0 , y0 ) denotes the position of the retinal image pixel under study. M (x, y) and O(x, y) refer to the value of the con-

199 200

4


Fig. 6. Peak images determined by a 2-D circular convolution mask shown in the upper left corner. (a) Peak image produced through the convolution of the gray orientation map in Fig. 5(a). (b) Peak image produced through the convolution of the binary orientation map in Fig. 5(b).

202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220

volution mask and the orientation map at (x, y), respectively. Parameter m denotes the radius of the circular convolution mask that can be similarly set as p. For the orientation maps in Fig. 5(a) and (b), Fig. 6(a) and (b) shows the determined peak images. As shown in Fig. 6, a peak is properly produced at the OD position. On the other hand, a peak is also produced at the macula center (i.e., fovea) that often has similar peak amplitude to the peak at the OD center. This can be explained by similar brightness variation structure around the macula, where the image variation along the line segment crossing the macula center reaches the maximum, whereas that along the orthogonal line segment [similar to Lc and Lt in Fig. 1(b)] reaches the minimum. The only difference is that the OD center is brighter than the surrounding pixels, whereas the macula center is darker. We, therefore, first classify the peaks into an OD category and a macula category, respectively. The classification is based on the image difference between the retinal image pixels at the peak center and those surrounding the peak center. The image difference is evaluated by two concentric circles as follows: Diff(x, y) =

221 222 223 224 225 226 227 228 229 230 231 232

where P (x, y) denotes thenormalized peak image. The symbol ∗ denotes dot product and Diff(x, y) > 0 sets all retinal image pixels with a negative image difference to zero. The OD can, therefore, be detected by the peak in the OD category that has the maximum score. For the example retinal image in Fig. 1(a), Fig. 7(a) shows the score image determined by the peak image in Fig. 6(b). It should be noted that the image difference is evaluated only at the detected peaks in practical implementation. The score image in Fig. 7(a) (as well as in Fig. 7(b), 9, and 10) where the image difference is evaluated at every pixel is just for the illustration purpose.

IE E Pr E oo f

201

Fig. 7. OD/Macula detection. (a) Score image by (9) for OD detection. (b) Score image by (10) for macula detection.

1 Ni

R1 d=0

I(d) −

1 No

R2

I(d)

(8)

d=0

where I refers to the retinal image under study and d denotes the distance between the peak and the surrounding retinal image pixels. R1 and R2 specify the radius of an inner concentric circle and an outer concentric circle where R2 is set at 2R1 . Ni and No denote the numbers of the retinal image pixels within the two concentric circles. In our system, we set R1 at (p − 1)/2, where p is the length of the line operator. The peak can, therefore, be classified to the OD or macula category, if the image difference is positive or negative, respectively. Finally, we detect the OD by a score that combines both the peak amplitude and the image intensity difference that by itself is also a strong indicator of the OD S(x, y) = P (x, y)(Diff(x, y) ∗ (Diff(x, y) > 0))

(9)

233 234 235 236 237 238 239 240 241 242 243

D. Discussion

244

It should be noted that though we build the orientation map by using the orientation of the line segment with the maximum variation, the orientation map can be built by the orientation of the line segment with the minimum variation with little effect on the OD detection performance. In addition, either the binary orientation map or the gray orientation map can be used to build the peak images with little effect on the OD detection performance either. Furthermore, the proposed line operator can be extended to locate the macula with little adaptation. With the determined peak image and the difference image, another score image can be similarly determined as follows:

245

S(x, y) = P (x, y) · (−Diff(x, y) ∗ (Diff(x, y) < 0))

246 247 248 249 250 251 252 253 254 255 256

(10)

where (Diff(x, y) < 0) sets all image pixels with a positive difference to zero and (−Diff(x, y)) reverses the value of image pixels having a negative difference. The macula can accordingly be located by the peak within the macula category that has the maximum score. For the retinal image in Fig. 1(a), Fig. 7(b) shows the score image determined by (10). As shown in Fig. 7(b), the peak with the maximum score is exactly located at the macula center.

257 258 259 260 261 262 263 264


Fig. 8. OD detection accuracy of the proposed technique in relation to the line length p and the convolution mask size m.

III. EXPERIMENTAL RESULTS

266

This section presents experimental results. Four public datasets used are first described. The performance of the designed line operator is then described and discussed.

267 268 269

A. Data Description

280

We evaluate our proposed technique by using four public datasets. In particular, the first two datasets are DIARETDB0 [26] and DIARETDB1 [27], which are composed of 130 and 89 retinal images and created for benchmarking diabetic retinopathy detection. The third dataset is DRIVE project’s dataset [28] that is composed of 40 retinal images and created for benchmarking retinal blood vessel extraction. The last one is STARE project’s dataset [16], which is composed of 50 images of pathological retina and 31 images of healthy retina. It is created for benchmarking OD detection and is much more challenging compared with the other three datasets.

281

B. OD Localization Results

282

For each retinal image within the four datasets, we first manually label 10–30 OD boundary pixels and then fit an OD boundary ellipse. The OD is deemed as located correctly, if the detected OD lies within the fitted boundary ellipse. Fig. 8 shows the average OD detection accuracy of the four public datasets. As shown in Fig. 8, the average OD detection accuracy varies within a small range when p and m change within a specific range (i.e., from R/10 to R/5), and a top average accuracy 97.4% (331 out of 340) is achieved when p and m are set at R/8. In addition, the top accuracies of the four datasets reach up to 99.2%, 98.9%, 97.5%, and 96.3%, respectively, when p and m vary between R/10 and R/5. In particular, most failed retinal images are among the 50 images of pathological retinas within STARE project’s dataset, many of which are severely degraded by different retinal lesion and imaging artifacts as shown in Figs. 9–11 and, therefore, do not have a clear OD-specific circular brightness structure. Furthermore, the OD detection accuracy drops when p and m become too large or too small. The accuracy drop can be explained by the fact that both

270 271 272 273 274 275 276 277 278 279

283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299

Q3 300

p and m are set based on OD size which usually varies within a specific range. Figs. 9 and 10 illustrate the OD detection results under the presence of retinal lesion and imaging artifacts. In particular, the three rows in the two figures show the test retinal images (detected OD is labeled by “+”), the derived orientation maps, and the final score images, respectively. As shown in Figs. 9 and 10, the line operator is able to detect the OD under the presence of retinal lesion, such as drusen (in the fifth image in Fig. 9), exudates (in the second and fourth images in Fig. 9), microaneurysms (in the fifth image in Fig. 10), papillary swelling (in the first image in Fig. 10), and hemorrhage (in the first image in Fig. 9), and imaging artifacts, such as haze (in the 2nd image in Fig. 10) and uneven illumination (in the third and fourth images in Fig. 10), that often produce regions with higher image brightness or image variation than the OD. Such results are due to the line operator that is specially designed to capture the OD-specific circular brightness structure. Table I compares the accuracies of the proposed technique and some earlier reported methods based on STARE project’s dataset. As shown in Table I, the proposed technique significantly outperforms the image-characteristics-based methods [11], [13] that cannot handle various types of imaging artifacts and retinal lesion properly. In addition, the accuracy of our proposed technique is close to that of the methods [15]–[18] that rely on the retinal blood vessels. As a comparison, the proposed technique requires no retinal blood vessels. In fact, all failed retinal images reported in [15] and [17] (i.e., the fourth image in Fig. 9 and the first and fourth images in Fig. 10) can be correctly detected by the proposed line operator. It should be noted that we only compare on STARE project’s dataset because STARE project’s dataset contains up to 50 images of pathological retinas and is widely used for benchmarking in the literature. Besides, many OD detection methods, including those based on the retinal blood vessels and our proposed method in this paper, are capable of detecting the OD from normal retinal images properly. In fact, all failed retinal images in STARE project’s dataset (by our proposed method) are from the 50 images of pathological retinas, and the remaining 31 normal ones are all correctly detected.

IE E Pr E oo f

265

5

301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340

C. Discussion

341

The designed line operator can be used for macula detection as described in Section II. We test the macula detection based on four subdatasets including 114, 85, 35, and 39 retinal images that are selected from the four public datasets. The use of four subdatasets is because of many retinal images in the four datasets, such as the third and fourth images in Fig. 9 and the first image in Fig. 10, do not have a discernible macula. Experiments over the four subdatasets show that an average macula detection accuracy of 98.2% is achieved. In addition, it takes around 40 s for our system to process a retinal image of original size (around 700 × 600 pixels). The detection speed could be improved significantly through code optimization and implementation in C. In addition, the designed line operator is robust against lower image resolution. We have

342 343 344 345 346 347 348 349 350 351 352 353 354 355

6


IE E Pr E oo f

Fig. 9. OD detection examples: The first row shows five retinal images within the four datasets that suffer from various types of imaging artifacts and retinal lesion (detected OD is labeled by “+”). The second and third rows show the corresponding binary orientation maps (p = R/7) and the score images, respectively.

Fig. 10. OD detection example. The first row shows five retinal images within the four datasets that suffer from different types of retinal lesion and imaging artifacts (detected OD is labeled by “+”). The second and third show the corresponding binary orientation maps (p = R/7) and the score images, respectively. 356 357 358 359 360 361 362 363 364 365 366 367 368

tested our system on half-sized retinal images (both p and m are half-sized accordingly) within the four public datasets. Experiments show that the optimal OD detection accuracy still reaches up to 95.9%, but the detection speed is improved tremendously by up to 12 times faster than that of retinal images of original size. Finally, the proposed technique still has several limitations. First, the proposed line operator is designed based on the assumption that the OD is more or less brighter than the surrounding retinal pixels and, therefore, cannot handle a very small number of retinal images whose OD is even darker than the surrounding pixels. Second, the proposed technique cannot handle the retinal images that do not have a clear circular

TABLE I COMPARISON OF THE OD DETECTION METHODS ON STARE PROJECT’S DATASET (THE ACCURACIES OF SINTHANAYOTHINA et al. [13] AND WALTER AND KLEIN [11] ARE BOTH TAKEN FROM HAAR [18])

brightness structure around their OD. Third, the performance of the proposed technique should be improved further through the incorporation of the anatomical relation between the OD and

369 370 371


375

D. Conclusion

376

383

This paper presents an automatic OD detection technique. A line operator is designed, which locates the OD through the detection of the OD-specific circular brightness structure. Compared with the reported techniques, the proposed technique requires neither the retinal blood vessel nor the macula. At the same time, it is tolerant to different types of retinal lesion and imaging artifacts. Experiments over four public datasets show that an accuracy of 97.4% is obtained.

384

REFERENCES

385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437

[1] K. Akita and H. Kuga, “A computer method of understanding ocular fundus images,” Pattern Recognit., vol. 15, no. 6, pp. 431–443, 1982. [2] N. Patton, T. M. Aslam, T. MacGillivary, I. J. Deary, B. Dhillon, R. H. Eikelboom, K. Yogesan, and I. J. Constable, “Retinal image analysis: Concepts, applications and potential,” Prog. Retin. Eye Res., vol. 25, no. 1, pp. 99–127, 2006. [3] T. Walter, J. C. Klein, P. Massin, and A. Erginay, “A Contribution of Image Processing to the Diagnosis of Diabetic Retinopathy-Detection of Exudates in Color Fundus Images of the Human Retina,” IEEE Trans. Med. Imag., vol. 21, no. 10, pp. 1236–1243, Oct. 2002. [4] R. Chrastek, M. Wolf, K. Donath, H. Niemann, D. Paulus, T. Hothorn, B. Lausen, R. Lammer, C. Y. Mardin, and G. Michelson, “Automated segmentation of the optic nerve head for diagnosis of glaucoma,” Med. Image Anal., vol. 9, no. 4, pp. 297–314, 2005. [5] A. D. Fleming, K. A. Goatman, S. Philip, J. A. Olson, and P. F. Sharp, “Automatic detection of retinal anatomy to assist diabetic retinopathy screening,” Phys. Med. Biol., vol. 52, no. 2, pp. 331–345, 2007. [6] A. Pinz, S. Bernogger, P. Datlinger, and A. Kruger, “Mapping the Human Retina,” IEEE Trans. Med. Imag., vol. 17, no. 4, pp. 606–619, Aug. 1998. [7] K. W. Tobin, E. Chaum, V. P. Govindasamy, and T. P. Karnowski, “Detection of anatomic structures in human retinal imagery,” IEEE Trans. Med. Imag., vol. 26, no. 12, pp. 1729–1739, Dec. 2007. [8] M. Niemeijer, M. D. Abramoff, and B. V. Ginneken, “Segmentation of the optic disc, macula and vascular arch in fundus photographs,” IEEE Trans. Med. Imag., vol. 26, no. 1, pp. 116–127, Jan. 2007. [9] W. Hsu, P. M. D. S. Pallawala, M. L. Lee, and K. A. Eong, “The Role of Domain Knowledge in the Detection of Retinal Hard Exudates,” in Proc. Int. Conf. Comp. Vis. Pattern Recognit., vol. 2, 2001, pp. 246–251. [10] Z. B. Sbeh, L. D. Cohen, G. Mimoun, and G. Coscas, “A new approach of geodesic reconstruction for drusen segmentation in eye fundus images,” IEEE Trans. Med. Imag., vol. 20, no. 12, pp. 1321–1333, Dec. 2002. [11] T. Walter and J. C. Klein, “Segmentation of color fundus images of the human retina: Detection of the optic disc and the vascular tree using morphological techniques,” in Proc. Int. Symp. Med. Data Anal., 2001, pp. 282–287. [12] H. Li and O. Chutatape, “Automatic location of optic disc in retinal images,” in Proc. Int. Conf. Image, vol. 2, pp. 837–840, 2001. [13] C. Sinthanayothina, J. F. Boycea, H. L. Cookb, and T. H. Williamsonb, “Automated localisation of the optic disc, fovea, and retinal blood vessels from digital colour fundus images,” Br. J. Ophthalmol., vol. 83, pp. 902– 910, 1999. [14] S. Sekhar, W. Al-Nuaimy, and A. K. Nandi, “Automated localisation of retinal optic disk using Hough transform,” in Proc. Int. Symp. Biomed. Imag.: Nano Macro, 2008, pp. 1577–1580. [15] A. Youssif, A. Z. Ghalwash, and A. Ghoneim, “Optic disc detection from normalized digital fundus images by means of a vessels’ direction matched filter,” IEEE Trans. Med. Imag., vol. 27, no. 1, pp. 11–18, Jan. 2008. [16] A. Hoover and M. Goldbaum, “Locating the optic nerve in a netinal image using the fuzzy convergence of the blood vessels,” IEEE Trans. Med. Imag., vol. 22, no. 8, pp. 951–958, Aug. 2003. [17] M. Foracchia, E. Grisan, and A. Ruggeri, “Detection of optic disc in retinal images by means of a geometrical model of vessel structure,” IEEE Trans. Med. Imag., vol. 23, no. 10, pp. 1189–1195, Oct. 2004.

373

377 378 379 380 381 382

[18] F. Haar, “Automatic localization of the optic disc in digital colour images of the human retina,” M.S. thesis, Utrecht University, Utrecht, The Netherlands, 2005. [19] H. Li and O. Chutatape, “Automated feature extraction in color retinal images by a model based approach,” IEEE Trans. Biomed. Eng., vol. 51, no. 2, pp. 246–254, Feb. 2004. [20] A. P. Rovira and E. Trucco, “Robust optic disc location via combination of weak detectors,” in Proc. Int. Conf. IEEE Eng. Med. Bio. Soc., 2008, pp. 3542–3545. [21] R. Zwiggelaar, S. M. Astley, C. R. M. Boggis, and C. J. Taylor, “Linear structures in mammographic images: Detection and classification,” IEEE Trans. Med. Imag., vol. 23, no. 9, pp. 1077–1086, Sep. 2004. [22] E. Ricci and R. Perfetti, “Retinal blood vessel segmentation using line operators and support vector classification,” IEEE Trans. Med. Imag., vol. 26, no. 10, pp. 1357–1365, Oct. 2007. [23] A. Osareh, M. Mirmehdi, B. Thomas, and R. Markham, “Comparison of colour spaces for optic disc localisation in retinal images,” in Proc. Int. Conf. Pattern Recognit., vol. 1, 2002, pp. 743–746. [24] C. Tomasi and R. Manduchi, “Bilateral Filtering for Gray and Color Images,” in Proc. IEEE Int. Conf. Comp. Vis., 1998, pp. 839–846. [25] W. Tasman and E. A. Jaeger, Duane’s Ophthalmology, 15th ed. Baltimore, MD: Lippincott Williams & Wilkins, 2009. [26] T. Kauppi, V. Kalesnykiene, J. K. Kamarainen, L. Lensu, I. Sorri, H. Uusitalo, H. Kälviäinen, and J. Pietilä, “DIARETDB0: Evaluation database and methodology for diabetic retinopathy algorithms,” Tech. Rep., Lappeenranta Univ. Technol., Lappeenranta, Finland, 2006. [27] T. Kauppi, V. Kalesnykiene, J. K. Kamarainen, L. Lensu, I. Sorri, H. Uusitalo, H. Klviinen, and J. Pietil, “DIARETDB1 diabetic retinopathy database and evaluation protocol,” Tech. Rep., Lappeenranta Univ. Technol., Lappeenranta, Finland, 2007. [28] J. J. Staal, M. D. Abramoff, M. Niemeijer, M. A. Viergever, and B. V. Ginneken, “Ridge based vessel segmentation in color images of the retina,” IEEE Trans. Med. Imag., vol. 23, no. 4, pp. 501–509, Apr. 2004.

IE E Pr E oo f

374

the macula, since the designed line operator is able to locate the macula with little adaptation. We will study these three issues in our future works.

372

7

438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470

Shijian Lu (M’xx) received the Ph.D. degree in electrical and computer engineering from National University of Singapore, Singapore, in 2005. He is currently a Senior Research Fellow at the Institute for Infocomm Research, Agency for Science, Technology and Research (A*STAR), Singapore. His research interests include document image analysis and medical image analysis. He has authored or coauthored more than 40 peer-reviewed journal and conference papers. Dr. Lu is a member of International Association of Pattern Recognition

471 472 473 Q4 474 475 476 477 478 479 480 481 482 483

Joo Hwee Lim (M’xx) received the B.Sc. and M.Sc. degrees in computer science from the National University of Singapore, Singapore, and the Ph.D. degree in computer science & engineering from the University of New South Wales, Sydney, Australia. Since October 1990, he has been with the Institute for Infocomm Research (I2R), Agency for Science, Technology and Research (A*STAR), Singapore, where he is currently the Head of the Computer Vision and Image Understanding Department. He is also an Adjunct Associate Professor at the School of Computer Engineering, Nanyang Technological University, Singapore. He is also the Co-Director of Image and Pervasive Access Laboratory (IPAL), a French–Singapore Joint Lab (UMI 2955) for the tenure January 2007 to December 2010, and the Director (Imaging) of a new joint lab (SAILOR) between I2R and Singapore Eye Research Institute for the tenure June 2009 to June 2012, where computer scientists and clinicians collaborate closely. He has authored or coauthored more than 170 international refereed journal and conference papers. He has also coauthored 16 patents (awarded and pending). His research interests include connectionist expert systems, neural-fuzzy systems, handwritten character recognition, multiagent systems, content-based image/video retrieval, scene/object recognition, medical image analysis. Dr. Lim was bestowed the title of “Chevallet dans l’ordre des Palmes Academiques” by the French Government in 2008.

484 Q5 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 Q6 506 507 508

(IAPR).

QUERIES

510

Q1: Author: Please spell out “DRIVE” and STARE in full, if possible. Q2. Author: There is discrepancy in the terms “c(ξ; x)” and “s(f (ξ), f (x))” between display equation and text. Please check and confirm. Q3. Author: The citation for Fig. 11 has been provided in text. However, there are only ten figures in the manuscript. Please check and confirm. Q4. Author: Please provide the year information in which S. Lu became a Member of IEEE. Q5. Author: Please provide the year information in which J. H. Lim became a Member of IEEE. Q6. Author: Please spell out “UMI” and “SAILOR” in full, if possible.

511 512 513 514 515 516 517

IE E Pr E oo f

509


Automatic Optic Disc Detection From Retinal Images by a Line Operator

1

2

Shijian Lu*, Member, IEEE, and Joo Hwee Lim, Member, IEEE

3

21 22

Abstract—Under the framework of computer-aided eye disease diagnosis, this paper presents an automatic optic disc (OD) detection technique. The proposed technique makes use of the unique circular brightness structure associated with the OD, i.e., the OD usually has a circular shape and is brighter than the surrounding pixels whose intensity becomes darker gradually with their distances from the OD center. A line operator is designed to capture such circular brightness structure, which evaluates the image brightness variation along multiple line segments of specific orientations that pass through each retinal image pixel. The orientation of the line segment with the minimum/maximum variation has specific pattern that can be used to locate the OD accurately. The proposed technique has been tested over four public datasets that include 130, 89, 40, and 81 images of healthy and pathological retinas, respectively. Experiments show that the designed line operator is tolerant to different types of retinal lesion and imaging artifacts, and an average OD detection accuracy of 97.4% is obtained. Index Terms—Computer-aided diagnosis, line operators, optic disc (OD) detection, retinal image analysis.

23

I. INTRODUCTION

24

UTOMATIC optic disc (OD) detection from retinal images is a very important task in ocular image analysis [1], [2] and computer-aided diagnosis of various types of eye diseases [3]–[5]. It is often a key step for the detection of other anatomical retinal structures, such as retinal blood vessels and macula [1], [6], [7], [8]. More importantly, it helps to establish a retinal coordinate system that can be used to determine the position of other retinal abnormalities, such as exudates, drusen, and hemorrhages [9], [10]. Some OD detection techniques have been reported in the literature. The early techniques make use of different types of ODspecific image characteristics. In particular, some techniques search for the brightest regions [11], [12] or regions with the highest image variation [13], [14] resulting from the bright OD and the dark blood vessels within the OD. The limitation of these methods is that many retinal images suffer from various types of

25 26 27 28 29 30 31 32 33 34 35 36 37 38 39

A

retinal lesion, such as drusen, exudates, microaneurysms, and hemorrhage, and imaging artifacts, such as haze, lashes, and uneven illumination (as illustrated in Figs. 9–11) that often produce brighter regions or regions with higher image variation compared with the OD. Several OD detection techniques make use of anatomical structures among the OD, macula, and retinal blood vessels. For example, some methods are based on the anatomical structure that all major retinal blood vessels radiate from the OD [15]–[18]. Some other methods make use of the relative position between the OD and the macula that often varies within a small range [19], [20]. Compared with the image characteristics, the anatomical structures are more reliable under the presence of retinal lesion and imaging artifacts. However, the extraction of either retinal blood vessels or the macula is often a nontrivial task by itself. This paper presents a line operator that is designed to locate the OD from retinal images accurately. Line operators have been used to locate linear structures from different types of images. For example, Zwiggelaar et al. used a line operator to detect linear structures from mammographic images [21], where a line strength is evaluated by the difference between the largest average image intensity along one oriented line segment and the average image intensity within a local neighborhood window. Ricci and Perfetti [22] used a similar line operator to detect the linear structures that are associated with the retinal blood vessels. Our proposed line operator is designed to capture the circular brightness structure associated with the OD. In particular, it evaluates the image variation along multiple oriented line segments and locates the OD based on the orientation of the line segment with the maximum/minimum variation. Fig. 1(a) shows an example of retinal image in DRIVE project’s dataset [16], and Fig. 1(b) shows an image that simulates the circular brightness structure associated with the OD. As shown in Fig. 1, the OD has a specific brightness variation pattern where the image variation along Lc in Fig. 1(b) across the OD center usually reaches the maximum, whereas that along Lt orthogonal to Lc reaches the minimum. The proposed method has several advantages. First, the designed line operator is tolerant to the retinal lesion and various types of imaging artifacts that most image-characteristics-based methods cannot handle properly. The tolerance to the imaging artifacts and retinal lesion can be explained by the proposed line operator that is designed to capture the unique circular brightness structure associated with the OD. Second, the designed line operator is stable and easy for implementation. It requires neither the retinal blood vessel nor the macula information. Third, the designed line operator can be extended for macula detection

IE E Pr E oo f

4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

1

Manuscript received March 16, 2010; revised July 22, 2010; accepted September 22, 2010. Date of publication; date of current version. Asterisk indicates corresponding author. ∗ S. Lu is with the Department of Computer Vision and Image Understanding, Institute for Infocomm Research, Agency for Science, Technology and Research (A*STAR), Singapore 138632 (e-mail: [email protected]). J. H. Lim is with the Department of Computer Vision and Image Understanding, Institute for Infocomm Research, Agency for Science, Technology and Research (A*STAR), Singapore 138632 (e-mail: [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TBME.2010.2086455

0018-9294/$26.00 © 2010 IEEE

40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88

2


Fig. 1. Circular brightness structure associated with the OD. (a) Example of retinal image in DRIVE project’s dataset with OD labeled by a bold black circle. (b) Simulated circular brightness structure: L c crossing the OD center and L t orthogonal to L c are added to illustrate the line segments along which the image variation reaches the maximum and the minimum.

Q1

Fig. 3. Example line operator that uses 20 oriented line segments and set the line length p at 21.

IE E Pr E oo f

smoothing filter [24] that combines geometric closeness and photometric similarity as follows: ∞ ∞ −1 h(x) = k (x) f (ξ)c(ξ, x)s(f (ξ); f (x))dξ (1) −∞

−∞

with the normalization factor ∞ ∞ k(x) = c(ξ, x)s(f (ξ); f (x))dξ −∞

Fig. 2. Retinal image preprocessing. (a) Lightness of the example retinal image in LAB color space. (b) Enhanced retinal image by bilateral smoothing where multiple crosses along a circle label the pixels to be used to illustrate the image variation along multiple oriented line segments.

89 90 91 92 93 94

with little adaptation. Experiments over four public datasets verify its superior performance. The rest of this paper is organized as follows. Section II describes the proposed OD detection technique. Experimental results are then described and discussed in Section III. Some concluding remarks are finally drawn in Section IV.

95

II. PROPOSED METHOD

96

This section presents the proposed OD detection technique. In particular, we divide this section into four subsections, which deal with the retinal image preprocessing, designed line operator, OD detection, and discussion, respectively.

97 98 99

100

A. Retinal Image Preprocessing

101

Retinal images need to be preprocessed before the OD detection. As the proposed technique makes use of the circular brightness structure of the OD, the lightness component of a retinal image is first extracted. We use the lightness component within the LAB color space, where the OD detection usually performs the best [23]. For the retinal image in Fig. 1(a), Fig. 2(a) shows the corresponding lightness image. The retinal image is then smoothed to enhance the circular brightness structure associated with the OD. We use a bilateral

102 103 104 105 106 107 108 109

110 111

112

(2)

−∞

where f (x) denotes the retinal image under study. c(ξ; x) and s(f (ξ), f (x)) measure the geometric closeness and the photometric similarity between the neighborhood center x and a nearby point ξ. We set both c(ξ; x) and s(f (ξ), f (x)) by Gaussian functions. The geometric spread σd and the photometric spread σr of the two Gaussian functions are typically set at 10 and 1 as reported in [24]. For the retinal image in Fig. 2(a), Fig. 2(b) shows the filtered retinal image.

113 114 115 116 117 Q2 118 119 120

B. Designed Line Operator

121

A line operator is designed to the detect circular regions that have similar brightness structure as the OD. For each image pixel at (x, y), the line operator first determines n line segments Li , i = 1, . . . , n of specific length p (i.e., number of pixels) at multiple specific orientations that center at (x, y). The image intensity along all oriented line segments can thus be denoted by a matrix I(x, y)n ×p , where each matrix row stores the intensity of p image pixels along one specific line segment. Fig. 3 shows an example of line operator that uses 20 oriented line segments and sets the line length p = 21. As shown in Fig. 3, each line segment Li at one specific orientation can be divided into two line segments Li,1 and Li,2 of the same length (p − 1)/2 by the image pixel under study (i.e., the black cell in Fig. 3). The image variation along the oriented line segments can be estimated as follows:

122

Di (x, y) = fm dn (IL i , 1 (x, y)) − fm dn (IL i , 2 (x, y)) , i = 1, . . . , n

(3)

123 124 125 126 127 128 129 130 131 132 133 134 135 136


Fig. 4. Image variation along multiple oriented line segments: Each graph shows the image variation vector D(x, y) of one retinal image pixel labeled by a cross along the circle in Fig. 2(b).

138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163

where fm dn (·) denotes a median function. fm dn (IL i , 1 (x, y)) and fm dn (IL i , 2 (x, y)) return the median image intensity along Li,1 and Li,2 , respectively. D = [D1 (x, y), . . . , Di (x, y), . . . , Dn (x, y)] is, therefore, a vector of dimension n that stores the image variations along n-oriented line segments. The orientation of the line segment with the maximum/minimum variation has specific pattern that can be used to locate the OD accurately. For retinal image pixels, which are far away from the OD, the orientation of the line segment with the maximum/minimum variation is usually arbitrary, but for those around the OD, the image variation along Lc [labeled in Fig. 1(b)] usually reach the maximum, whereas that along Lt reaches the minimum. Fig. 4 shows the image variation vectors D(x, y) of eight pixels that are labeled by crosses along a circle shown in Fig. 2(b). Suppose that there is a Cartesian coordinate system centered at the OD, as shown in Fig. 2(b). For the retinal image pixels in quadrants I and III, the image variations along the 1st–10th [i.e., Lt in Fig. 1(b)] and the 11th–20th (i.e., Lc ) line segments labeled in Fig. 3 reach the minimum and the maximum, respectively, as shown in Fig. 4. But for the retinal image pixels in quadrants II and IV, the image variations along the 1st–10th and the 11th–20th line segments instead reach the maximum and the minimum, respectively. An orientation map can, therefore, be constructed based on the orientation of the line segment with the maximum (or minimum) variation as follows: O(x, y) = argmax D(x, y)

(4)

i

164 165 166 167

Fig. 5. Orientation map of the retinal image in Fig. 2(b). (a) Gray orientation map that is determined by using (4). (b) Binary orientation map that is determined by using (5).

where n refers to the number of the oriented line segments used in the line operator. For the retinal image in Fig. 1(a), Fig. 5(a) and (b) shows the determined gray orientation map and binary orientation map, respectively. As shown in Fig. 5(a), for retinal image pixels in quadrants I and III around the OD, the orientation map is a bit dark because the orientation of the line segment with the maximum variation usually lies between 1 and (n/2) + 1. However, for retinal image pixels in quadrants II and IV, the orientation map is bright because the orientation of the line segment with the maximum variation usually lies between n/2 and n. The binary orientation map in Fig. 5(b) further verifies such orientation pattern. The OD will then be located by using the orientation map to be described in the following.

IE E Pr E oo f

137

3

where D(x, y) denotes the image variation vector evaluated in (3). In addition, a binary orientation map can also be constructed by classifying the orientation of the line segment with the maximum variation into two categories as follows: Q(x, y) =

−1, 1,

if argmax D(x, y) < i

otherwise

n +1 2

168 169 170 171 172 173 174 175 176 177 178 179 180 181

C. OD Detection

182

We use a line operator of 20 oriented line segments because line operators with more line segments have little effect on the orientation map. The line length p is set as follows:

183

p = kR

(6)

where R denote the radius of the central circular region of retinal images as illustrated in Fig. 1(a). Parameter k controls the line length, which usually lies between 1/10 and 1/5 based on the relative OD size within retinal images [25]. The use of R incorporates possible variations of the image resolution. The specific pattern within the orientation map is captured by a 2-D circular convolution mask shown at the upper left corner of two peak images in Fig. 6. As shown in Fig. 6, the convolution mask can be divided into four quadrants, where the cells within quadrants I and III are set at −1, whereas those within quadrants II and IV are set at 1 based on the specific pattern within the orientation map. An orientation map can thus be converted into a peak image as follows: P (x, y) =

x 0 +m

y 0 +m

184 185

M (x, y)O(x, y)

186 187 188 189 190 191 192 193 194 195 196 197 198

(7)

x=x 0 −m y =y 0 −m

(5)

where (x0 , y0 ) denotes the position of the retinal image pixel under study. M (x, y) and O(x, y) refer to the value of the con-

199 200

4


Fig. 6. Peak images determined by a 2-D circular convolution mask shown in the upper left corner. (a) Peak image produced through the convolution of the gray orientation map in Fig. 5(a). (b) Peak image produced through the convolution of the binary orientation map in Fig. 5(b).

202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220

volution mask and the orientation map at (x, y), respectively. Parameter m denotes the radius of the circular convolution mask that can be similarly set as p. For the orientation maps in Fig. 5(a) and (b), Fig. 6(a) and (b) shows the determined peak images. As shown in Fig. 6, a peak is properly produced at the OD position. On the other hand, a peak is also produced at the macula center (i.e., fovea) that often has similar peak amplitude to the peak at the OD center. This can be explained by similar brightness variation structure around the macula, where the image variation along the line segment crossing the macula center reaches the maximum, whereas that along the orthogonal line segment [similar to Lc and Lt in Fig. 1(b)] reaches the minimum. The only difference is that the OD center is brighter than the surrounding pixels, whereas the macula center is darker. We, therefore, first classify the peaks into an OD category and a macula category, respectively. The classification is based on the image difference between the retinal image pixels at the peak center and those surrounding the peak center. The image difference is evaluated by two concentric circles as follows: R1 R2 1 1 Diff(x, y) = I(d) − I(d) Ni No d=0

221 222 223 224 225 226 227 228 229 230 231 232

where P (x, y) denotes thenormalized peak image. The symbol ∗ denotes dot product and Diff(x, y) > 0 sets all retinal image pixels with a negative image difference to zero. The OD can, therefore, be detected by the peak in the OD category that has the maximum score. For the example retinal image in Fig. 1(a), Fig. 7(a) shows the score image determined by the peak image in Fig. 6(b). It should be noted that the image difference is evaluated only at the detected peaks in practical implementation. The score image in Fig. 7(a) (as well as in Fig. 7(b), 9, and 10) where the image difference is evaluated at every pixel is just for the illustration purpose.

IE E Pr E oo f

201

Fig. 7. OD/Macula detection. (a) Score image by (9) for OD detection. (b) Score image by (10) for macula detection.

(8)

d=0

where I refers to the retinal image under study and d denotes the distance between the peak and the surrounding retinal image pixels. R1 and R2 specify the radius of an inner concentric circle and an outer concentric circle where R2 is set at 2R1 . Ni and No denote the numbers of the retinal image pixels within the two concentric circles. In our system, we set R1 at (p − 1)/2, where p is the length of the line operator. The peak can, therefore, be classified to the OD or macula category, if the image difference is positive or negative, respectively. Finally, we detect the OD by a score that combines both the peak amplitude and the image intensity difference that by itself is also a strong indicator of the OD S(x, y) = P (x, y)(Diff(x, y) ∗ (Diff(x, y) > 0))

(9)

233 234 235 236 237 238 239 240 241 242 243

D. Discussion

244

It should be noted that though we build the orientation map by using the orientation of the line segment with the maximum variation, the orientation map can be built by the orientation of the line segment with the minimum variation with little effect on the OD detection performance. In addition, either the binary orientation map or the gray orientation map can be used to build the peak images with little effect on the OD detection performance either. Furthermore, the proposed line operator can be extended to locate the macula with little adaptation. With the determined peak image and the difference image, another score image can be similarly determined as follows:

245

S(x, y) = P (x, y) · (−Diff(x, y) ∗ (Diff(x, y) < 0))

246 247 248 249 250 251 252 253 254 255 256

(10)

where (Diff(x, y) < 0) sets all image pixels with a positive difference to zero and (−Diff(x, y)) reverses the value of image pixels having a negative difference. The macula can accordingly be located by the peak within the macula category that has the maximum score. For the retinal image in Fig. 1(a), Fig. 7(b) shows the score image determined by (10). As shown in Fig. 7(b), the peak with the maximum score is exactly located at the macula center.

257 258 259 260 261 262 263 264


Fig. 8. OD detection accuracy of the proposed technique in relation to the line length p and the convolution mask size m.

III. EXPERIMENTAL RESULTS

266

This section presents experimental results. Four public datasets used are first described. The performance of the designed line operator is then described and discussed.

267 268 269

A. Data Description

280

We evaluate our proposed technique by using four public datasets. In particular, the first two datasets are DIARETDB0 [26] and DIARETDB1 [27], which are composed of 130 and 89 retinal images and created for benchmarking diabetic retinopathy detection. The third dataset is DRIVE project’s dataset [28] that is composed of 40 retinal images and created for benchmarking retinal blood vessel extraction. The last one is STARE project’s dataset [16], which is composed of 50 images of pathological retina and 31 images of healthy retina. It is created for benchmarking OD detection and is much more challenging compared with the other three datasets.

281

B. OD Localization Results

282

For each retinal image within the four datasets, we first manually label 10–30 OD boundary pixels and then fit an OD boundary ellipse. The OD is deemed as located correctly, if the detected OD lies within the fitted boundary ellipse. Fig. 8 shows the average OD detection accuracy of the four public datasets. As shown in Fig. 8, the average OD detection accuracy varies within a small range when p and m change within a specific range (i.e., from R/10 to R/5), and a top average accuracy 97.4% (331 out of 340) is achieved when p and m are set at R/8. In addition, the top accuracies of the four datasets reach up to 99.2%, 98.9%, 97.5%, and 96.3%, respectively, when p and m vary between R/10 and R/5. In particular, most failed retinal images are among the 50 images of pathological retinas within STARE project’s dataset, many of which are severely degraded by different retinal lesion and imaging artifacts as shown in Figs. 9–11 and, therefore, do not have a clear OD-specific circular brightness structure. Furthermore, the OD detection accuracy drops when p and m become too large or too small. The accuracy drop can be explained by the fact that both

270 271 272 273 274 275 276 277 278 279

283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299

Q3 300

p and m are set based on OD size which usually varies within a specific range. Figs. 9 and 10 illustrate the OD detection results under the presence of retinal lesion and imaging artifacts. In particular, the three rows in the two figures show the test retinal images (detected OD is labeled by “+”), the derived orientation maps, and the final score images, respectively. As shown in Figs. 9 and 10, the line operator is able to detect the OD under the presence of retinal lesion, such as drusen (in the fifth image in Fig. 9), exudates (in the second and fourth images in Fig. 9), microaneurysms (in the fifth image in Fig. 10), papillary swelling (in the first image in Fig. 10), and hemorrhage (in the first image in Fig. 9), and imaging artifacts, such as haze (in the 2nd image in Fig. 10) and uneven illumination (in the third and fourth images in Fig. 10), that often produce regions with higher image brightness or image variation than the OD. Such results are due to the line operator that is specially designed to capture the OD-specific circular brightness structure. Table I compares the accuracies of the proposed technique and some earlier reported methods based on STARE project’s dataset. As shown in Table I, the proposed technique significantly outperforms the image-characteristics-based methods [11], [13] that cannot handle various types of imaging artifacts and retinal lesion properly. In addition, the accuracy of our proposed technique is close to that of the methods [15]–[18] that rely on the retinal blood vessels. As a comparison, the proposed technique requires no retinal blood vessels. In fact, all failed retinal images reported in [15] and [17] (i.e., the fourth image in Fig. 9 and the first and fourth images in Fig. 10) can be correctly detected by the proposed line operator. It should be noted that we only compare on STARE project’s dataset because STARE project’s dataset contains up to 50 images of pathological retinas and is widely used for benchmarking in the literature. Besides, many OD detection methods, including those based on the retinal blood vessels and our proposed method in this paper, are capable of detecting the OD from normal retinal images properly. In fact, all failed retinal images in STARE project’s dataset (by our proposed method) are from the 50 images of pathological retinas, and the remaining 31 normal ones are all correctly detected.

IE E Pr E oo f

265

5

301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340

C. Discussion

341

The designed line operator can be used for macula detection as described in Section II. We test the macula detection based on four subdatasets including 114, 85, 35, and 39 retinal images that are selected from the four public datasets. The use of four subdatasets is because of many retinal images in the four datasets, such as the third and fourth images in Fig. 9 and the first image in Fig. 10, do not have a discernible macula. Experiments over the four subdatasets show that an average macula detection accuracy of 98.2% is achieved. In addition, it takes around 40 s for our system to process a retinal image of original size (around 700 × 600 pixels). The detection speed could be improved significantly through code optimization and implementation in C. In addition, the designed line operator is robust against lower image resolution. We have

342 343 344 345 346 347 348 349 350 351 352 353 354 355

6


IE E Pr E oo f

Fig. 9. OD detection examples: The first row shows five retinal images within the four datasets that suffer from various types of imaging artifacts and retinal lesion (detected OD is labeled by “+”). The second and third rows show the corresponding binary orientation maps (p = R/7) and the score images, respectively.

Fig. 10. OD detection example. The first row shows five retinal images within the four datasets that suffer from different types of retinal lesion and imaging artifacts (detected OD is labeled by “+”). The second and third show the corresponding binary orientation maps (p = R/7) and the score images, respectively. 356 357 358 359 360 361 362 363 364 365 366 367 368

tested our system on half-sized retinal images (both p and m are half-sized accordingly) within the four public datasets. Experiments show that the optimal OD detection accuracy still reaches up to 95.9%, but the detection speed is improved tremendously by up to 12 times faster than that of retinal images of original size. Finally, the proposed technique still has several limitations. First, the proposed line operator is designed based on the assumption that the OD is more or less brighter than the surrounding retinal pixels and, therefore, cannot handle a very small number of retinal images whose OD is even darker than the surrounding pixels. Second, the proposed technique cannot handle the retinal images that do not have a clear circular

TABLE I COMPARISON OF THE OD DETECTION METHODS ON STARE PROJECT’S DATASET (THE ACCURACIES OF SINTHANAYOTHINA et al. [13] AND WALTER AND KLEIN [11] ARE BOTH TAKEN FROM HAAR [18])

brightness structure around their OD. Third, the performance of the proposed technique should be improved further through the incorporation of the anatomical relation between the OD and

369 370 371


375

D. Conclusion

376

383

This paper presents an automatic OD detection technique. A line operator is designed, which locates the OD through the detection of the OD-specific circular brightness structure. Compared with the reported techniques, the proposed technique requires neither the retinal blood vessel nor the macula. At the same time, it is tolerant to different types of retinal lesion and imaging artifacts. Experiments over four public datasets show that an accuracy of 97.4% is obtained.

384

REFERENCES

385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437

[1] K. Akita and H. Kuga, “A computer method of understanding ocular fundus images,” Pattern Recognit., vol. 15, no. 6, pp. 431–443, 1982. [2] N. Patton, T. M. Aslam, T. MacGillivary, I. J. Deary, B. Dhillon, R. H. Eikelboom, K. Yogesan, and I. J. Constable, “Retinal image analysis: Concepts, applications and potential,” Prog. Retin. Eye Res., vol. 25, no. 1, pp. 99–127, 2006. [3] T. Walter, J. C. Klein, P. Massin, and A. Erginay, “A Contribution of Image Processing to the Diagnosis of Diabetic Retinopathy-Detection of Exudates in Color Fundus Images of the Human Retina,” IEEE Trans. Med. Imag., vol. 21, no. 10, pp. 1236–1243, Oct. 2002. [4] R. Chrastek, M. Wolf, K. Donath, H. Niemann, D. Paulus, T. Hothorn, B. Lausen, R. Lammer, C. Y. Mardin, and G. Michelson, “Automated segmentation of the optic nerve head for diagnosis of glaucoma,” Med. Image Anal., vol. 9, no. 4, pp. 297–314, 2005. [5] A. D. Fleming, K. A. Goatman, S. Philip, J. A. Olson, and P. F. Sharp, “Automatic detection of retinal anatomy to assist diabetic retinopathy screening,” Phys. Med. Biol., vol. 52, no. 2, pp. 331–345, 2007. [6] A. Pinz, S. Bernogger, P. Datlinger, and A. Kruger, “Mapping the Human Retina,” IEEE Trans. Med. Imag., vol. 17, no. 4, pp. 606–619, Aug. 1998. [7] K. W. Tobin, E. Chaum, V. P. Govindasamy, and T. P. Karnowski, “Detection of anatomic structures in human retinal imagery,” IEEE Trans. Med. Imag., vol. 26, no. 12, pp. 1729–1739, Dec. 2007. [8] M. Niemeijer, M. D. Abramoff, and B. V. Ginneken, “Segmentation of the optic disc, macula and vascular arch in fundus photographs,” IEEE Trans. Med. Imag., vol. 26, no. 1, pp. 116–127, Jan. 2007. [9] W. Hsu, P. M. D. S. Pallawala, M. L. Lee, and K. A. Eong, “The Role of Domain Knowledge in the Detection of Retinal Hard Exudates,” in Proc. Int. Conf. Comp. Vis. Pattern Recognit., vol. 2, 2001, pp. 246–251. [10] Z. B. Sbeh, L. D. Cohen, G. Mimoun, and G. Coscas, “A new approach of geodesic reconstruction for drusen segmentation in eye fundus images,” IEEE Trans. Med. Imag., vol. 20, no. 12, pp. 1321–1333, Dec. 2002. [11] T. Walter and J. C. Klein, “Segmentation of color fundus images of the human retina: Detection of the optic disc and the vascular tree using morphological techniques,” in Proc. Int. Symp. Med. Data Anal., 2001, pp. 282–287. [12] H. Li and O. Chutatape, “Automatic location of optic disc in retinal images,” in Proc. Int. Conf. Image, vol. 2, pp. 837–840, 2001. [13] C. Sinthanayothina, J. F. Boycea, H. L. Cookb, and T. H. Williamsonb, “Automated localisation of the optic disc, fovea, and retinal blood vessels from digital colour fundus images,” Br. J. Ophthalmol., vol. 83, pp. 902– 910, 1999. [14] S. Sekhar, W. Al-Nuaimy, and A. K. Nandi, “Automated localisation of retinal optic disk using Hough transform,” in Proc. Int. Symp. Biomed. Imag.: Nano Macro, 2008, pp. 1577–1580. [15] A. Youssif, A. Z. Ghalwash, and A. Ghoneim, “Optic disc detection from normalized digital fundus images by means of a vessels’ direction matched filter,” IEEE Trans. Med. Imag., vol. 27, no. 1, pp. 11–18, Jan. 2008. [16] A. Hoover and M. Goldbaum, “Locating the optic nerve in a netinal image using the fuzzy convergence of the blood vessels,” IEEE Trans. Med. Imag., vol. 22, no. 8, pp. 951–958, Aug. 2003. [17] M. Foracchia, E. Grisan, and A. Ruggeri, “Detection of optic disc in retinal images by means of a geometrical model of vessel structure,” IEEE Trans. Med. Imag., vol. 23, no. 10, pp. 1189–1195, Oct. 2004.

373

377 378 379 380 381 382

[18] F. Haar, “Automatic localization of the optic disc in digital colour images of the human retina,” M.S. thesis, Utrecht University, Utrecht, The Netherlands, 2005. [19] H. Li and O. Chutatape, “Automated feature extraction in color retinal images by a model based approach,” IEEE Trans. Biomed. Eng., vol. 51, no. 2, pp. 246–254, Feb. 2004. [20] A. P. Rovira and E. Trucco, “Robust optic disc location via combination of weak detectors,” in Proc. Int. Conf. IEEE Eng. Med. Bio. Soc., 2008, pp. 3542–3545. [21] R. Zwiggelaar, S. M. Astley, C. R. M. Boggis, and C. J. Taylor, “Linear structures in mammographic images: Detection and classification,” IEEE Trans. Med. Imag., vol. 23, no. 9, pp. 1077–1086, Sep. 2004. [22] E. Ricci and R. Perfetti, “Retinal blood vessel segmentation using line operators and support vector classification,” IEEE Trans. Med. Imag., vol. 26, no. 10, pp. 1357–1365, Oct. 2007. [23] A. Osareh, M. Mirmehdi, B. Thomas, and R. Markham, “Comparison of colour spaces for optic disc localisation in retinal images,” in Proc. Int. Conf. Pattern Recognit., vol. 1, 2002, pp. 743–746. [24] C. Tomasi and R. Manduchi, “Bilateral Filtering for Gray and Color Images,” in Proc. IEEE Int. Conf. Comp. Vis., 1998, pp. 839–846. [25] W. Tasman and E. A. Jaeger, Duane’s Ophthalmology, 15th ed. Baltimore, MD: Lippincott Williams & Wilkins, 2009. [26] T. Kauppi, V. Kalesnykiene, J. K. Kamarainen, L. Lensu, I. Sorri, H. Uusitalo, H. Kälviäinen, and J. Pietilä, “DIARETDB0: Evaluation database and methodology for diabetic retinopathy algorithms,” Tech. Rep., Lappeenranta Univ. Technol., Lappeenranta, Finland, 2006. [27] T. Kauppi, V. Kalesnykiene, J. K. Kamarainen, L. Lensu, I. Sorri, H. Uusitalo, H. Klviinen, and J. Pietil, “DIARETDB1 diabetic retinopathy database and evaluation protocol,” Tech. Rep., Lappeenranta Univ. Technol., Lappeenranta, Finland, 2007. [28] J. J. Staal, M. D. Abramoff, M. Niemeijer, M. A. Viergever, and B. V. Ginneken, “Ridge based vessel segmentation in color images of the retina,” IEEE Trans. Med. Imag., vol. 23, no. 4, pp. 501–509, Apr. 2004.

IE E Pr E oo f

374

the macula, since the designed line operator is able to locate the macula with little adaptation. We will study these three issues in our future works.

372

7

438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470

Shijian Lu (M’xx) received the Ph.D. degree in electrical and computer engineering from National University of Singapore, Singapore, in 2005. He is currently a Senior Research Fellow at the Institute for Infocomm Research, Agency for Science, Technology and Research (A*STAR), Singapore. His research interests include document image analysis and medical image analysis. He has authored or coauthored more than 40 peer-reviewed journal and conference papers. Dr. Lu is a member of International Association of Pattern Recognition

471 472 473 Q4 474 475 476 477 478 479 480 481 482 483

Joo Hwee Lim (M’xx) received the B.Sc. and M.Sc. degrees in computer science from the National University of Singapore, Singapore, and the Ph.D. degree in computer science & engineering from the University of New South Wales, Sydney, Australia. Since October 1990, he has been with the Institute for Infocomm Research (I2R), Agency for Science, Technology and Research (A*STAR), Singapore, where he is currently the Head of the Computer Vision and Image Understanding Department. He is also an Adjunct Associate Professor at the School of Computer Engineering, Nanyang Technological University, Singapore. He is also the Co-Director of Image and Pervasive Access Laboratory (IPAL), a French–Singapore Joint Lab (UMI 2955) for the tenure January 2007 to December 2010, and the Director (Imaging) of a new joint lab (SAILOR) between I2R and Singapore Eye Research Institute for the tenure June 2009 to June 2012, where computer scientists and clinicians collaborate closely. He has authored or coauthored more than 170 international refereed journal and conference papers. He has also coauthored 16 patents (awarded and pending). His research interests include connectionist expert systems, neural-fuzzy systems, handwritten character recognition, multiagent systems, content-based image/video retrieval, scene/object recognition, medical image analysis. Dr. Lim was bestowed the title of “Chevallet dans l’ordre des Palmes Academiques” by the French Government in 2008.

484 Q5 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 Q6 506 507 508

(IAPR).

QUERIES

510

Q1: Author: Please spell out “DRIVE” and STARE in full, if possible. Q2. Author: There is discrepancy in the terms “c(ξ; x)” and “s(f (ξ), f (x))” between display equation and text. Please check and confirm. Q3. Author: The citation for Fig. 11 has been provided in text. However, there are only ten figures in the manuscript. Please check and confirm. Q4. Author: Please provide the year information in which S. Lu became a Member of IEEE. Q5. Author: Please provide the year information in which J. H. Lim became a Member of IEEE. Q6. Author: Please spell out “UMI” and “SAILOR” in full, if possible.

511 512 513 514 515 516 517

IE E Pr E oo f

509