Automated feature extraction for early detection of diabetic retinopathy i

000
001
002
003
004
005
006
007
008
009
010
011
012
013
014
015
016
017
018
019
020
021
022
023
024
025
026
027
028
029
030
031
032
033
034
035
036
037
038
039
040
041
042
043
044
045
046
047
048
049
050
051
052
053
054
055
056
057
058
059
060
061
062
063
064
065
066
067
068
069
070
071
072
073
074
075
076
077
078
079
080
081
082
083
084
085
086
087
088
089
090
091
092
093
094
095
096
097
098
099
100
101
102
103
104
105
106
107
CVPR
#****
CVPR
#****
CVPR 2009 Submission #****. CONFIDENTIAL REVIEW COPY. DO NOT DISTRIBUTE.
Automated Feature Extraction for Early Detection of Diabetic Retinopathy in
Fundus Images
Anonymous CVPR submission
Paper ID ****
Abstract
Automated detection of lesions in retinal images can as-
sist in early diagnosis and screening of a common dis-
ease:Diabetic Retinopathy. A robust and computationally
efficient approach for the localization of the different fea-
tures and lesions in a fundus retinal image is presented
in this paper. Since many features have common intensity
properties, geometric features and correlations are used to
distinguish between them. We propose a new constraint for
optic disk detection where we first detect the major blood
vessels first and use the intersection of these to find the
approximate location of the optic disk. This is further lo-
calized using color properties. We also show that many
of the features such as the blood vessels, exudates and mi-
croaneurysms and hemorrhages can be detected quite ac-
curately using different morphological operations applied
appropriately. Extensive evaluation of the algorithm on
a database of 516 images with varied contrast, illumina-
tion and disease stages yields 97.1% success rate for optic
disk localization, a sensitivity and specificity of 95.7% and
94.2% respectively for exudate detection and 95.1% and
90.5% for microaneurysm/hemorrhage detection. These
compare very favorably with existing systems and promise
real deployment of these systems.
1. Introduction
Diabetic retinopathy (DR) is a common retinal compli-
cation associated with diabetes. It is a major cause of blind-
ness in both middle and advanced age groups. According
to the National Diabetes Information data (US) 1
, a total
of 20.8 million people i.e. 7 percent of the US population
have diabetes out of which only 14.6 million cases are diag-
nosed. Early detection of the disease via regular screening
is particularly important to prevent vision loss. In this pro-
cess, an automated DR diagnostic system can assist in a big
1http://www.diabetes.niddk.nih.gov/dm/pubs/statistics
Figure 1. Illustration of various features on a typical retionopathic
image.
way since a large population has to be screened and that too
repeatedly.
Color fundus images are used by ophthalmologists to
study eye diseases like diabetic retinopathy. Figure 1 shows
a typical retinal image labeled with various feature compo-
nents of Diabetic Retinopathy. Microaneurysms are small
saccular pouches caused by local distension of capillary
walls and appear as small red dots[1]. This may also lead
to big blood clots called hemorrhages. Hard exudates are
yellow lipid deposits which appear as bright yellow lesions.
The bright circular region from where the blood vessels em-
anate is called the optic disk. The fovea defines the center
of the retina, and is the region of highest visual acuity. The
spatial distribution of exudates and microaneurysms and
hemorrhages, especially in relation to the fovea can be used
to determine the severity of diabetic retinopathy.
1.1. Related Work
Sinthaniyothin [12] uses maximum variance to obtain
the optic disk center and a region growing segmentation
method to obtain the exudates. [4] tracks the optic disk
through a pyramidal decomposition and obtains disk local-
ization from a template-based matching that uses the Haus-
dorff distance measure on the binary edge image. However,
the above methods will fail if exudates similar in brightness
and size to the optic disk are present.

108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
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
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
CVPR
#****
CVPR
#****
[1, 7] used blood vessel intersection property to obtain
the optic disk. However, they use the whole blood vessel
network which can lead to wrong or inconclusive results
because of noise from the fringe blood vessels. In contrast,
we use only the main blood vessels, which is more robust.
Statistical classification techniques have been very pop-
ular lately for the problem of lesion classification. Exudates
have color properties similar to the optic disk while Microa-
neurysms are difficult to segment due to their similarity in
color and proximity with blood vessels. In order to classify
detected features, typically, candidate regions are detected
using color/morphological techniques and then classifica-
tion is done on these regions using some classifier. Many
classifiers have been tried including Neural Networks [12],
PCA [9], Fuzzy C-means clustering [10], SVMs ([16],[2],
[15]) and simple Bayesian classification ([15], [13]).
STARE is a complete system for various retinal diseases
[6]. The optic disk is detected using blood vessel conver-
gence and high intensity property. In order to determine
the features and classification method to be used for a given
lesion, a Bayesian probabilistic system is used.
In this paper, we develop methods to automatically de-
tect all of these features in a fundus image using image
processing techniques. We show that many of the features
such as the blood vessels, exudates and microaneurysms
and hemorrhages can be detected quite accurately using
different morphological operations applied appropriately.
Blood vessels of different thicknesses can be extracted us-
ing open and close operations. Exudates appear as bright
patches with sharp edges in retinal images and can be ex-
tracted using open and close operations using filters of dif-
ferent sizes. Microaneurysms and Hemorrhages (MAHMs)
are segmented using morphological filters that exploit their
local ’dark patch’ property. These are further classified as
lesion/non-lesion using a color model extracted from the
blood vessels. We propose a new constraint for optic disk
detection where we first detect the major blood vessels and
then use the intersection of these to find the approximate
location of the optic disk. This is further localized using
color properties. Detection of the Optic disk, fovea and the
blood vessels is used not only for distinguishing them from
lesions but also for extracting color information for better
lesion detection.
The rest of the paper is organized as follows: Sec-
tion 2 details the blood vessel extraction algorithm while
section 3 elucidates the exudate detection procedure; Sec-
tion 4 describes the optic disk detection method and sec-
tion 5 presents the method for detection of fovea, microa-
neurysms and hemorrhages. Section 6 describes determin-
ing the severity of the disease using lesion detection while
in section 7, the results of the algorithm over an extensive
dataset are presented.
2. Multi-Scale Blood Vessel Extraction
In our approach, color images input from the fundus
camera are initially resized to a standard size of 768 × 576
pixels while maintaining the original aspect ratio. We se-
lect the green channel for all our operations because retinal
images are almost always saturated in the red channel and
have very low contrast in the blue channel.
A closing operation is performed on the green channel
image using two different sizes of a structuring element (fil-
ter). Closing operation is defined as dilation (Max filter) fol-
lowed by erosion (Min filter). The formulations of dilation
and erosion for gray scale images are as follows.
Dilation:
A ⊕ B = A1(x, y) = max
i,j∈B
(A(x − i, y − j) + B(i, j))
Erosion:
A ⊖ B = A2(x, y) = min
i,j∈B1
(A(x − i, y − j) + B1(i, j))
where A is the input image, B and B1 are the structur-
ing elements or masks used for dilation and erosion respec-
tively.
Dilation in gray scale enlarges brighter regions and
closes small dark regions. The erosion is necessary to shrink
the dilated objects back to their original size and shape.
The dark regions closed by dilation do not respond to ero-
sion. Thus, the vessels being thin dark segments laid out on
a brighter background are closed by such a closing opera-
tion. A subtraction of the closed images across two different
scales (let S1 and S2 be the sizes of the structuring elements
B1 and B2) will thus give the blood vessel segments of the
green channel image. The operation is as follows:
C′
= (A ⊕ B2) ⊖ B2 − (A ⊕ B1) ⊖ B1
We use a disk shaped structuring element for morphological
operations. The radius of the larger disk (S2) is fixed at
a high value (we use 6 pixels for an image of size 768 ×
576 pixels) so that all the vessels including the main blood
vessel get closed. S1 is chosen adaptively as follows:
1. 1 or 2 pixels below S2 if we want to obtain only the
thicker vessels emanating from the optic disk.
2. Atleast 4 pixels below S2 to obtain the entire blood
vessel network.
Criterion 1 is used for optic disk localization whereas cri-
terion 2 is used in microaneurysms and hemorrhages de-
tection. The image C′
is thresholded (90% of the max-
imum intensity) and median filtered to obtain the binary
image of the blood vessels (U). Morphological thinning is
then performed on U to obtain the skeleton of the blood
2

216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
CVPR
#****
CVPR
#****
(a) (b)
(c) (d)
Figure 2. Illustration of the multi-scale vessel extraction algorithm
on an image of our dataset: (a) The original green channel image,
(b) The green image dilated at scale S2, (c) The vessel network
obtained at lower scale S1 (criterion 2), (d) The thicker vessels
obtained using higher S1 (criterion 1).
vessel network. Thinning operation is implemented as
U − (U ⊖ B1 − U ⊖ B2), where B1 and B2 are disjoint
structuring elements and U is the complement of the image
U. Noise can occur in the thinned image usually in the form
of dots. A 2 × 2 median filtering operation is performed to
remove the isolated specks of noise. The vessel segments
being connected structures are unaffected by this operation.
An additional source of noise in retinopathic images could
be exudates, the removal of which is detailed in section 6.
Figure 2 shows the results of the vessel extraction algo-
rithm on an image having no exudates. Figure 2 (c) show-
ing the entire vessel network was obtained using criterion
2 while Figure 2 (d) having only the thicker vessels arising
from the optic disk was obtained using criterion 1.
3. Exudate Localization and Detection
Exudates appear as bright lesions in retinopathic im-
ages and have sharp edges and high contrast with the back-
ground. Most of the standard edge detectors like Sobel and
Canny add a lot of noise and miss out key edges when used
for extracting exudate edges and hence are not suitable for
this application. We perform boundary detection for exu-
dates using morphological operations.
Dilation is performed on the green channel (obtained af-
ter pre-processing without equalization and normalization)
at 2 different scales: S3 and S4, both of which are greater
than S2 which was used for vessel extraction. Hence, at
both S3 and S4, the blood vessels do not appear in the di-
lated result. The exudates being bright with sharp edges
respond to dilation. Subtraction of the results across the 2
scales gives the boundaries of the exudates: P = (A⊕B4)−
(A ⊕ B3). The image P is thresholded in intensity to obtain
the binary boundaries. The threshold is chosen as α times
the maximum intensity in P where α is obtained by training.
Hard exudates give closed boundaries in the thresholded re-
sult. Short breaks in the contours are connected by smooth-
ing splines. This bridging of short breaks in boundaries is
useful for extracting softer exudates. A morphological fill-
ing operation is then used to search for regions bounded by
closed contours in the result. It is defined as follows:
Ek
= (Ek−1
⊕ B) ∩ Hc
(1)
This is iterated starting from k = 1 until Ek
= Ek−1
. Here,
Hc
is the complement of the thresholded binary image, B
is a 4-connected structuring element and E0
is an image
containing a seed point inside the boundary. In the case of
multiple holes, this operation is done in each hole until all
the holes are filled.
Morphological filling operation on the binary image thus
gives us the candidate exudate patches. However, the candi-
date regions may contain artifacts. Therefore, a linear clas-
sifier is built which uses the brightness and edge properties
of exudates. Exudates are bright yellow or white in color
and have high intensity in the green channel. We localize
the exudate patches more accurately by taking all the candi-
date regions whose mean intensities in the green channel are
greater than a fraction β (obtained by training) of the max-
imum intensity in the channel. For classifying the patches
based on their edge strength, the gradient magnitude image
of the green channel is chosen. This gradient magnitude
image is thresholded (the absolute threshold γ obtained by
training) and the number of white pixels in the thresholded
image for each exudate patch is counted. We denote this as
the gradient count of each patch. Patches which do not have
sufficient gradient count (δ) are discarded.
Patches that satisfy both the brightness criterion and gra-
dient count are retained. In each of the patches classified as
exudates, the exact lesion boundary is tracked starting from
the pixel with the highest gradient magnitude and complet-
ing the contour based on continuity of gradient magnitude
and direction. Pixels in the interior of these contours are
then accurately classified as exudate pixels. The optic disk
which may invariably appear in the result is masked out us-
ing the procedure outlined in the next section.
The various thresholds like α, β etc. were obtained by
training the exudate algorithm on a set of 50 color fundus
images obtained from ophthalmologists. Ground truth im-
ages were obtained for this set with exudates pre-marked
and the parameter values that gave optimal results (in terms
of detecting true exudates and not detecting artifacts) were
chosen. The values we obtained are α = 0.06, β = 0.52,
3

324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
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
CVPR
#****
CVPR
#****
(a) (b)
(c)
Figure 3. Illustration of the exudate algorithm on an image of our
dataset: (a) The original image, (b) Localized candidate regions,
(c) Final result with exudate pixels in black color and optic disk
masked out.
γ = 3 and δ = 4.
Figure 3 shows the steps of the exudate detection algo-
rithm on an image of our dataset. The candidate regions ob-
tained after morphological segmentation are shown in Fig-
ure 3 (b) and the result after classification, pixel identifica-
tion and optic disk removal is shown in Figure 3 (c).
4. Optic Disk Detection
The detection of the optic disk in fundus images is a very
important task because of its similarity in brightness, color
and contrast to the exudates. It invariably appears in exudate
detection results and hence there is a need to mask it out.
Moreover, the optic disk is an important retinal feature and
can be used for registration of retinal images. It can also be
used to diagnose other diseases like Glaucoma.
In this paper, we use a constraint that has been over-
looked so far which is to detect the optic disk using con-
vergence of only the thicker blood vessels arising from it.
This significantly improves the performance compared to
existing techniques that use the entire vessel network. Also,
this thicker vessel convergence is almost always present in
the image as opposed to other features of the optic disk such
as color or circular shape. This approach is combined with
the high intensity property of disk regions in a cost function
to improve the robustness of optic disk detection compared
to existing methods.
4.1. Exudate Subtraction
The vessel extraction algorithm detailed in section 3 (us-
ing criterion 1) yields the skeleton of the thicker blood ves-
sels. These ’thicker’ vessels include the main blood vessel
and other smaller but thick vessels emanating from the op-
tic disk. Exudates have the potential of occurring as noise
in the vessel segmentation because closing and subtract-
ing the irregularly shaped exudates using regular structur-
ing elements is not an exact operation (i.e. the shape of
exudates does not remain the same after the closing oper-
ation). Hence, we perform an additional step to overcome
this limitation. The result of the exudate detection algo-
rithm is dilated and subtracted from the blood vessel result.
This removes any noise due to exudates in the extraction of
the main blood vessel. The resulting image after exudate
subtraction is then processed to obtain the optic disk center.
4.2. Vessel Convergence
The segments of the thicker blood vessel skeleton
are modeled as lines. We transform the vessel image
into the Hough space using the Hough Transform (HT)
(x, y) HT
−−−→
(r, θ) to obtain the lines. The dataset of lines
thus generated is reduced by eliminating lines with slopes
θ < 450
. This can be done as the vessels converging at the
optic disk typically have a high slope in the vicinity of the
disk. This reduced dataset of lines is intersected pairwise
to generate an intersection map. Lines close to each other
and with nearly the same slope are not intersected due to a
higher triangulation error.
Weighted Convergence Optimization: The intersection
map generated from candidate line segments of the thicker
vessels is used to find the location of the optic disk. The
map is dilated to make the region of convergence more ap-
parent.
Dilation and erosion in binary images are implemented
as OR and AND filters (extension of Max and Min filters in
gray scale). Let the dilated intersection image be M and the
pre-processed green channel image be A. Then, we define
a weighted image (J) as follows:
J = M + wA
where w is obtained as follows:
w =
1, if N ≤ N0
(N0/N)β
, if N > N0
(2)
where N is the number of high intensity pixels
(Intensity > 200) in the initial green channel image (be-
fore pre-processing), N0 is the number of pixels corre-
sponding to the size of a normal disk (taken from [12]), β is
a power law factor (taken as 2) which rapidly decreases w
4

432
433
434
435
436
437
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
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
CVPR
#****
CVPR
#****
(a) (b) (c)
(d) (e) (f)
Figure 4. Illustration of the steps of the optic disk algorithm on an
image of our dataset: (a) The green channel image, (b) The major
vessels obtained using criterion 1, (c) The dilated intersection map,
(d) The image J obtained by weighted summation, (e) Final disk
center result indicated as a black spot, (f) The disk mask obtained.
as N increases above N0. Using this formula, the image A
is given less weightage if the number of bright pixels (N)
in the green channel is more than a threshold N0 (sign of
bright exudates); else A is given a weightage of 1 (normal
images). Image J is then used to obtain the optimal location
of the optic disk using a cost function F.
F =
i,j∈W
J(x − i, y − j) (3)
where (x, y) is a point in the intersection map and sum-
mation is over a circular sliding window (W). F can be
computed efficiently using dynamic programming/integral
images. The point (x, y) which maximizes F, is taken as
the location of the optic disk.
Figure 4 shows the steps of the optic disk localization
algorithm on an image of our dataset. The vessels of Figure
4 (b) are obtained after exudate subtraction.
5. Detection of Microaneurysms and Hemor-
rhages
Microaneurysms are small blood clots which occur due
to capillary burst. They are the hardest to detect in retino-
pathic images. Hemorrhages are bigger clots. Microa-
neurysms And HeMorrhages (MAHMs) are treated as holes
(i.e. small dark blobs sorrounded by brighter regions) and
morphological filling is performed on the green channel
(without pre-processing) to identify them. The filling op-
eration in gray scale is an extension of binary filling used
in section 5. The unfilled green channel image is then
subtracted from the filled one and thresholded in intensity
to yield an image (R) with microaneurysm patches. The
threshold (ν) is chosen based on the mean intensity of the
retinal image in the red channel. For a mean intensity of
127 (taken as reference on the intensity scale of 0-255),
ν was chosen as 7 (obtained by training). The thresh-
old is incremented/decremented by 1 for every 20 units in-
crease/decrease in the mean.
Blood vessels can also appear as noise in the microa-
neurysm and hemorrhage detection as they have similar
color and contrast to the clots. To remove this additional
noise, the full blood vessel network skeleton (section 4, Cri-
terion 2) is first obtained. The resulting blood vessel net-
work is dilated and subtracted from the image R to remove
the noise due to vessels. The remaining patches are further
classified using intensity properties and a color model based
on the detected blood vessels.
5.1. Blood Vessel based Color Model
Microaneurysms and hemorrhages have similar color
properties as the blood vessels. Hence, a color model is
built for classification of these lesions. We use the blood
vessel result obtained before the thinning operation is per-
formed (i.e. image U in section 4). For every candidate mi-
croaneurysm/hemorrhage patch, the blood vessel segments
in a local neighborhood are obtained. The mean (µ) and
standard deviation (σ) of these vessels in the red and green
channels are calculated. The pixels in each localized patch
whose intensities in the red and green channels are in the
range [µ − 1.2σ , µ + 1.2σ] are retained. This helps in the
removal of background artifacts.
5.2. Detection and Removal of Fovea
The fovea is a dark region located in the center of the
macula region of the retina. It commonly appears in mi-
croaneurysm and hemorrhage detection results much as the
optic disk does in exudate detection results. The fovea is
detected using the location of the optic disk and curvature
of the main blood vessel. The main blood vessel is obtained
as the thickest and largest blood vessel emanating from the
optic disk. The entire course of the main blood vessel is
obtained (from the image of the thicker vessels) by looking
for its continuity from the optic disk. This vessel is mod-
eled as a parabola ([9]). The vertex of the parabola is taken
as the pixel on the main blood vessel that is closest to the
center of the optic disk circular mask. The fovea is located
approximately between 2 to 3 optical disk diameter (ODD)
distance from the vertex, along the main axis of the modeled
parabola and is taken as the darkest pixel in this region. The
region of the fovea is taken to be 1 optic disk diameter of the
detected fovea location and detection of microaneurysms in
this region is suppressed.
Figure 5 shows the steps of the MAHMs algorithm on an
image of our dataset. We use a size threshold to distinguish
between microaneurysm and hemorrhage patches.
5

540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
CVPR
#****
CVPR
#****
(a) (b)
(c) (d)
Figure 5. Illustration of the steps of our Microaneurysms and Hem-
orrhages detection algorithm: (a) An image from our dataset, (b)
The image R obtained after morphological filling and threshold-
ing, and removing blood vessels and pixels not satisfying the color
model, (c) Fovea detection marked with ’+’, the main blood vessel
and optic disk mask boundary are also shown, (d) The final result
after removing fovea and classification, MAHMs marked in blue.
6. Predicting the Severity of Disease
The distribution of the exudate lesions about the fovea
can be used to predict the severity of Diabetic macular
edema. [9] has divided the retinal image into 10 sub-regions
about the fovea. The exudates occurring in the macular re-
gion are more dangerous and require immediate medical
attention than the ones farther away. Similarly, the size,
count and distribution of microaneurysms and hemorrhages
is also used to predict the severity of DR. The region around
the optic disk is divided into four quadrants for this pur-
pose. The International Council of Ophthalmology 2
lists
5 levels for Diabetic Retinopathy based on these criteria:
none, mild, moderate, severe, and proliferative. Our system
uses these criteria in order to classify each image in these
categories. Figure 5 contains a very large hemorrhage and
many other smaller hemorrhages and microaneurysms and
is a case of proliferative Diabetic Retinopathy.
7. Results of Experiments and Discussion
We used a dataset of 516 images for evaluating the algo-
rithm. The images were obtained from diverse sources and
hence have sufficient variations in color, illumination and
quality. The various sources of the images are as follows:
211 images were obtained from ophthalmologists at 2 lo-
cal eye hospitals; 81 images were taken from the STARE
2http://www.icoph.org/standards/pdrclass.html
Table 1. Results of optic disk localization for specific databases
and for the overall normal, abnormal cases.
Source or Type No. Images No. Correct % Success
Hospitals 211 209 99.1
STARE 81 76 93.8
DRIVE 40 40 100
Diaretdb0 130 127 97.7
Red Atlas 54 49 90.7
Overall Normal 112 111 99.1
Overall Abnormal 404 390 96.5
database [7]; 130 images were from the Diaretdb0 database
[8]; 40 images were from the DRIVE database [14] and 54
were from the Red Atlas database 3
.
The images in the dataset were classified by ophthal-
mologists based on the lesion type (exudates/MAHMs) into
those with the lesion and those without it. An image having
no lesions is considered normal whereas one that has lesions
like exudates, microaneurysms and hemorrhages is consid-
ered abnormal. Among the 211 images obtained from eye
hospitals, 29 are normal and 182 are abnormal. The DRIVE
database contains 33 normal and 7 abnormal images. The
STARE database has 30 images which are normal and the
other 51 being abnormal. The diaretdb0 database consisting
of 130 images has 20 normal and 110 abnormal ones. All
the images taken from the Red Atlas database were abnor-
mal.
Of the 516 images, 345 were identified by ophthalmolo-
gists as having exudates and 171 did not have any exudates.
A total of 348 images were identified as containing microa-
neurysms/hemorrhages while 168 were free from this le-
sion type. The entire algorithm was run on the database
and results for optic disk localization, exudate detection and
MAHMs detection were obtained. The matlab code takes
20 seconds per image on an average to run on a 2 GHz ma-
chine with 448 MB RAM.
Table 1 summarizes the results of our optic disk detec-
tion algorithm on the dataset. The pixel location obtained
prior to placing the circular mask is considered for evalu-
ation purposes. If the pixel location indicated by the optic
disk algorithm falls within the boundaries of the optic disk,
then it is considered a correct detection and vice-versa. The
results in Table 1 are classified based on the source and also
based on the nature of the images (normal or abnormal).
In Table 2, the overall optic disk detection rate of our al-
gorithm is presented along with other results from literature.
Gagnon et al. [4] obtain 100% optic disk detection but their
dataset consists of only normal images without retinopathy.
Adam Hoover et. al. [7] evaluate their optic disk detection
algorithm exclusively on the STARE database. We obtain
better results on the STARE database as indicated by table
3http://redatlas.org/main.htm
6

648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
CVPR
#****
CVPR
#****
Table 2. Comparison of results for optic disk localization
Author No. of Images % Success
Chutatape et al. [9] 35 99
Gagnon et al. [4] 40 100
Adam Hoover et al. [7] 81 89
Our Method 516 97.1
(a) (b)
(c) (d)
Figure 6. Optic disk localization (without mask) results indicated
as black spot on some images of the dataset: (a) Image with big,
bright exudates and bright optic disk, (b) An image with non uni-
form illumination from the diaretdb0 database, (c) An image with
lots of blood clots from the STARE database, (d) Another image
from the STARE database.
1. Figure 6 shows the optic disk localization results (prior
to the masking step) on some images of the dataset. Figures
6 (c) and 6 (d) show 2 cases where our algorithm performs
better than [7] on the STARE database.
We obtain sensitivity and specificity results for our al-
gorithm at both the image and pixel levels. For the pixel
level results, we obtained ground truth images from the oph-
thalmologists with lesion (both exudates and MAHMs) pix-
els marked for all images except those from the diaretdb0
database which already has the ground truth. The results of
the algorithm were evaluated with respect to the groundtruth
results to determine pixel level sensitivity and sepecificity
for each lesion type. The pixel level results were averaged
across all images.
In Table 3, the sensitivity and specificity results of our
exudate detection algorithm are presented for the 516 im-
ages along with other results from the literature. In Table 4,
the results for our microaneurysm and hemorrhage detec-
tion method are presented along with other results from the
Table 3. Comparison of Results for exudate detection: NI-Number
of Images used, NIEx-Number of Images with Exudates, SN-
Sensitivity, SP-Specificity (ND*-data not given in the paper)
Author NI NIEx SN(%) SP(%)
Chutatape [9] 35 28 100 71.0
Sinthanayothin [12] 30 21 88.5 99.7
Sanchez et al.[11] 20 10 100 90.0
Wang et al.[15] 154 54 100 70.0
Garcia et al.[5](image) 50 25 100 84.0
Garcia et al.[5](lesion) 50 25 84.4 62.7
Fleming et al.[3] 13219 300 95.0 84.6
Sopharak et al.[13] 10 ND* 93.38 98.14
Our approach(image) 516 345 95.7 94.2
Our approach(pixel) 516 345 94.6 91.1
Table 4. Comparison of Results for MAHMs Detection: NI-
Number of Images used, NIMAHMs-Number of Images with
MAHMs, SN-Sensitivity, SP-Specificity
Author NI NIMAHMs SN(%) SP(%)
Sinthanayothin [12] 30 14 77.5 88.7
Our result (Image) 516 348 95.1 90.5
Our result (Pixel) 516 348 92.0 90.1
(a) (b)
(c)
Figure 7. Our exudate and MAHMs detection results (indicated
in black and blue color respectively) on an image of the dataset
with the optic disk removed using a white mask in the exudate
result: (a) The original Image with hard exudates and many mi-
croaneurysms, (b) Our exudate detection result, (c) Our MAHMs
detection result (the result has been dilated to make the MAHM
patches visible).
literature. Figure 7 shows the exudate and MAHM detec-
tion results on an image of our dataset. Image 7 (a) contains
hard exudates and lots of microaneurysms/hemorrhages
which are detected by our algorithm as shown in Figures
7

756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
CVPR
#****
CVPR
#****
Figure 8. Plot of ROC curves for various lesion types: (a)The ROC
curve for exudates, (b) The ROC curve for MAHMs.
7 (b) and 7 (c) respectively. Some of the exudates in Figure
7 (b) are clustered in the macula and hence need immediate
medical attention. Based on the size, count and distribution
of MAHMs, our algorithm correctly predicted the level of
severity of diabetic retinopathy for this case as moderate.
Figures 8 (a) and 8 (b) show the ROC curves of our al-
gorithm for exudates and MAHMs respectively. The curves
are obtained by varying various parameter thresholds (α, β,
γ, δ for exudates and ν, ζ for MAHMs respectively) while
running the algorithm on the database and plotting the True
Positive rate (TP i.e. sensitivity) versus the False Positive
rate (FP i.e 1-True Negative rate). Here, the TP and FP are
obtained at the image level for each lesion type. The true
positive rates reach close to 100% for both exudates and
MAHMs at very low false positive rates, thus demonstrat-
ing the robustness of our algorithm for lesion detection.
8. Conclusion
In this paper, an efficient framework for early detection
of Diabetic Retinopathy has been developed. The optic disk
is tracked by combining the blood vessel convergence and
high disk intensity properties in a cost function. We show
that, as opposed to most methods that use learning tech-
niques, geometrical relationships of different features and
lesions can be used along with simple morphological oper-
ations in order to obtain a very robust system for analysis
of retinal images. Our techniques may further be combined
with some learning methods for possibly even better results.
References
[1] K. Akita and H. Kuga. A computer method of understand-
ing ocular fundus images. Pattern Recognition, 15(6):431–443,
1982. 1, 2
[2] O. Chutatape and X. Zhang. Detection and classification of
bright lesions in color fundus images. In International Confer-
ence on Image Processing, pages 139–142, 2004. 2
[3] A. D. Fleming, S. Philip, K. A. Goatman, G. J. Williams, J. A.
Olson, and P. F. Sharp. Automated detection of exudates for di-
abetic retinopathy screening. Physics in Medicine and Biology,
52:7385–7396, Dec. 2007. 7
[4] L. Gagnon, M. Lalonde, M. Beaulieu, and M.-C. Boucher.
Procedure to detect anatomical structures in optical fundus im-
ages. In Proc. SPIE Medical Imaging: Image Processing, pages
1218–1225, 2001. 1, 6, 7
[5] M. Garcia, R. Hornero, C. Sanchez, M. Lopez, and A. Diez.
Feature extraction and selection for the automatic detection of
hard exudates in retinal images. Engineering in Medicine and
Biology Society, 2007. EMBS 2007. 29th Annual International
Conference of the IEEE, pages 4969–4972, Aug. 2007. 7
[6] M. Goldbaum, S. Moezzi, A. Taylor, S. Chatterjee, J. Boyd,
E. Hunter, and R. Jain. Automated diagnosis and image under-
standing with object extraction, object classification, and infer-
encing in retinal images. International Conference on Image
Processing, 3:695–698, Sept. 1996. 2
[7] A. Hoover and M. GoldBaum. Locating the optic nerve in a
retinal image using the fuzzy convergence of the blood vessels.
IEEE Trans. on Medical Imaging, 22:951–958, Aug. 2003. 2, 6,
7
[8] T. Kauppi, V. Kalesnykiene, J. K. Kamarainen, L. Lensu,
I. Sorri, H. Uusitalo, H. Kalviainen, and J. Pietila. Diaretdb0:
Evaluation database and methodology for diabetic retinopathy
algorithms. Technical report, Lappeenranta University of Tech-
nology, Lappeenranta, Finland, 2006. 6
[9] H. Li and O. Chutatape. A model-based approach for auto-
mated feature extraction in fundus images. In Proc. IEEE Inter-
national Conference on Computer Vision, page 394, 2003. 2, 5,
6, 7
[10] A. Osareh. Automated Identification of Diabetic Retinal Ex-
udates and the Optic Disc. PhD thesis, Univ. of Bristol, Jan.
2004. 2
[11] C. I. Sanchez, A. Mayo, M. Garcia, M. I. Lopez, and
R. Hornero. Automatic image processing algorithm to de-
tect hard exudates based on mixture models. Engineering in
Medicine and Biology Society, 2006. EMBS ’06. 28th Annual
International Conference of the IEEE, pages 4453–4456, Sept.
2006. 7
[12] C. Sinthanayothin, J. F. Boyce, T. H. Williamson, H. L.
Cook, E. Mensah, S. Lal, and D. Usher. Automated detec-
tion of diabetic retinopathy on digital fundus images. Diabetic
Medicine, 19:105–112, 2002. 1, 2, 4, 7
[13] A. Sopharak, K. Thet Nwe, Y. A. Moe, M. N. Dailey, and
B. Uyyanonvara. Automatic exudate detection with a naive
bayes classifier. In International Conference on Embedded Sys-
tems and Intelligent Technology (ICESIT), pages 139–142, Feb.
2008. 2, 7
[14] J. Staal, M. D. Abramoff, M. Niemeijer, M. A. Viergever, and
B. V. Ginneken. Ridge-based vessel segmentation in color im-
ages of the retina. IEEE Trans. on Medical Imaging, 23(4):501–
509, 2004. 6
[15] H. Wang, W. Hsu, G. K. G., and L. M. L. An effective ap-
proach to detect lesions in color retinal images. In In Proc. Conf.
Comp. Vision Pattern Rec., pages II: 181–186, 2000. 2, 7
[16] X. Zhang and O. Chutatape. Top-down and bottom-up strate-
gies in lesion detection of background diabetic retinopathy. In
In Proc. Conf. Comp. Vision Pattern Rec., pages 422–428, 2005.
2
8

Automated feature extraction for early detection of diabetic retinopathy i

More Related Content

What's hot (20)

Viewers also liked (17)

Similar to Automated feature extraction for early detection of diabetic retinopathy i (20)

Automated feature extraction for early detection of diabetic retinopathy i