Improvement of automatic hemorrhages detection methods using brightness correction on fundus images

Abstract

We have been developing several automated methods for detecting abnormalities in fundus images. The purpose of this study is to improve our automated hemorrhage detection method to help diagnose diabetic retinopathy. We propose a new method for preprocessing and false positive elimination in the present study. The brightness of the fundus image was changed by the nonlinear curve with brightness values of the hue saturation value (HSV) space. In order to emphasize brown regions, gamma correction was performed on each red, green, and blue-bit image. Subsequently, the histograms of each red, blue, and blue-bit image were extended. After that, the hemorrhage candidates were detected. The brown regions indicated hemorrhages and blood vessels and their candidates were detected using density analysis. We removed the large candidates such as blood vessels. Finally, false positives were removed by using a 45-feature analysis. To evaluate the new method for the detection of hemorrhages, we examined 125 fundus images, including 35 images with hemorrhages and 90 normal images. The sensitivity and specificity for the detection of abnormal cases was were 80% and 88%, respectively. These results indicate that the new method may effectively improve the performance of our computer-aided diagnosis system for hemorrhages.

Full text

Page 1

*hatanaka@gifu-nct.ac.jp; phone 81 58 320 1384; fax 81 58 320 1263; gifu-nct.ac.jp
Improvement of Automatic Hemorrhages Detection Methods using
Brightness Correction on Fundus Images

Yuji Hatanaka*a, Toshiaki Nakagawa*b, Yoshinori Hayashi*c, Masakatsu Kakogawa*c,
Akira Sawada*d, Kazuhide Kawase*d, Takeshi Hara*b, Hiroshi Fujita*b
a Dept. of Electronic Control Engineering, Gifu National College of Technology
2236-2 Kamimakuwa, Motosu-shi, Gifu 501-0495, Japan
b Dept. of Intelligent Image Information, Division of Regeneration and Advanced Med. Science
Graduate School of Medicine, Gifu University, 1-1 Yanagido, Gifu-shi, Gifu 501-1194, Japan
c Tak Co., Ltd., 4-32-12 Kono, Ogaki-shi, Gifu 503-0803, Japan
d Dept. of Ophthalmology, Gifu University, School of Medicine
1-1 Yanagido, Gifu-shi, Gifu 501-1194, Japan


ABSTRACT

We have been developing several automated methods for detecting abnormalities in fundus images. The purpose of this
study is to improve our automated hemorrhage detection method to help diagnose diabetic retinopathy. We propose a
new method for preprocessing and false positive elimination in the present study. The brightness of the fundus image
was changed by the nonlinear curve with brightness values of the hue saturation value (HSV) space. In order to
emphasize brown regions, gamma correction was performed on each red, green, and blue-bit image. Subsequently, the
histograms of each red, blue, and blue-bit image were extended. After that, the hemorrhage candidates were detected.
The brown regions indicated hemorrhages and blood vessels and their candidates were detected using density analysis.
We removed the large candidates such as blood vessels. Finally, false positives were removed by using a 45-feature
analysis. To evaluate the new method for the detection of hemorrhages, we examined 125 fundus images, including 35
images with hemorrhages and 90 normal images. The sensitivity and specificity for the detection of abnormal cases was
were 80% and 88%, respectively. These results indicate that the new method may effectively improve the performance
of our computer-aided diagnosis system for hemorrhages.

Keywords: pre-processing, detection, retinal fundus image, diabetic retinopathy, hemorrhage


1. INTRODUCTION

The detection of hemorrhages is one of the important factors in the early diagnosis of diabetic retinopathy (DR). The
existence of hemorrhages is generally used to diagnose DR or hypertensive retinopathy by using the classification
scheme of Scheie. In spite of detecting microaneurysms, it is difficult for ophthalmologists to find them in noncontrast
fundus images. The contrast observed in a microaneurysm image is very low; therefore, ophthalmologists usually detect
microaneurysms by using fluorescein angiograms. However, it is difficult to use fluorescein as a contrast medium for
diagnosing all the medical examinees subjected to mass screening. Therefore, the patients who show the possibility of
having DR were thoroughly examined at a hospital in Japan.
The number of patients with adult diseases such as diabetes and hypertension is on the rise in Japan. To prevent or
detect these diseases in early stages, ophthalmologists rely on the examination of fundus images obtained from patients
aged over 40 years during complete health examinations or mass screenings. DR is a complication associated with
diabetes, and there is a high probability that diabetic patients will develop this condition within 10 years from the onset
of diabetes. Furthermore, DR is the leading cause of blindness. In Japan, there are approximately 7.4 million patients
with diabetes and approximately 16.2 million patients who may have diabetes [1]. Approximately three million are
thought to suffer from DR. This disease can be prevented from developing into blindness if it is treated at an early stage.
Medical Imaging 2008: Computer-Aided Diagnosis, edited by Maryellen L. Giger, Nico Karssemeijer
Proc. of SPIE Vol. 6915, 69153E, (2008) · 1605-7422/08/$18 · doi: 10.1117/12.771051
Proc. of SPIE Vol. 6915 69153E-1

Page 2

However, it has been recorded that approximately 3,000 people have lost their vision following the onset of DR. Fundus
photographs obtained by the fundus camera are used to diagnose DR. Japanese ophthalmologists usually examine the
presence of hemorrhages, microaneurysms, and exudates in order to diagnose DR.
Recently, many studies have been reported on the use of fundus images in the detection DR [2–9]. Niemeijer et al.
proposed a method for the detection of microaneurysms in fluorescein angiograms [2–4]. Their method comprised two
steps. Firstly, microaneurysms were detected by using a watershed transform. The false positives were then reduced by
reliable vessel detection in the vicinity of each microaneurysms candidate. Serrano et al. proposed a method for
detecting microaneurysms by the region growing technique in order to analyze fluorescein angiograms [5]. Usher et al.
presented a method for detecting hemorrhages, microaneurysms, and exudates [6, 7]. The region growing method used
by them segmented the abnormal and retinal regions. Nagayoshi et al. presented a method that included two additional
processes to this method [8]. One of these was the normalization of the color pixel values and the other was the use of
the color pixel values for blood vessels. Moreover, the previous method had the drawback of a long operation time of
approximately 43 s per image [8].
We also reported a method for detecting hemorrhages and exudates in noncontrast fundus images [9]. Although the
sensitivity for hemorrhages was 85%, the specificity was 21%. In our previous study, we had two problems—the
absence of a technique to normalize fundus images and the removal of false positives. In this study, we aim to develop
methods for fundus image normalization and false positive elimination method in the noncontrast images.


2. METHODS

2.1 Overall Scheme
The flowchart for our overall detection scheme is shown in Fig. 1. It consists of seven stages: (1) image digitization,
(2) image normalization, (3) extraction of optic nerve head, (4) detection of hemorrhage candidates, (5) elimination of
false positives in blood vessels, (6) elimination of funicular-shaped false positives, and (7) elimination of false positives
by feature analysis. Further details are described below.

2.2 Image digitization
One hundred forty five fundus images were captured using a fundus camera and a flatbed-type scanner. Eighty seven
fundus images were obtained at a resolution of an array of 1,600 × 1,600 pixels with 24-bit color, and 58 fundus images
were obtained at a resolution of an array of 2,800 × 2,800 pixels with 24-bit color. An example of a color fundus image
is shown in Fig. 2 (a).
Subsequently, the scale of the matrix was first reduced to the VGA size (width 640 pixels) by obtaining detailed
subsamples from the original image data to improve processing efficiency.

2.3 Image normalization
Due to the flash, there is an atypical change in the color of the fundus images. We suggested a scheme of brightness
correction using hue saturation value (HSV) space. First, the brightness values of the HSV space were calculated. The
brightness correction value Bc(i, j) is given by the following equation:

( , )1 { ( , ) 1} Bc i jV i j
=−

where V(i, j) is the brightness value of the HSV space.

( , ),V i jMAXMAX R i j G i j B i j
=

Next, the red value R(i, j), green value G(i, j), and blue value B(i, j) changed by Bc(i, j) are given by the following
equation:

2
−
(1)
: ( , ), ( , ), ( , )
(2)
Proc. of SPIE Vol. 6915 69153E-2

Page 3

Image digitization
Fundus image normalization
Extraction of optic nerve head
Detection of hemorrhage candidates
Elimination of blood vessel candidates
Elimination of funicular-shaped false positives
Elimination of
false positives on red-
bit image by rule-based
method
Elimination of
false positives on red-
bit image by
discemment machine
Elimination of
false positives on blue-
bit image by rule-based
method
Elimination of
false positives on blue-
bit image
by discemment machine
Elimination of
false positives on
green-bit image by
rule-based method
Elimination of
false positives on
green-bit image by
discemment machine
Results output


Fig. 1. Flowchart for detecting hemorrhages on fundus images.

( , )
60
( , )
60
( , )[1
( , )[1
( , ){1 [1
( , ) 0
( , ) 1
H i j
( , ) 2
( , ) 3
( , ) 4
i
H i j
( , ) 5
i
H i j
( , )mod6
( , )
f i j
( , )
( , )
( , )
( , )t i j
if
if
if
if
if
i
f
( , )]
( , ) ( , )]
( , )] ( , )}
( , )R i j
R
→
( , )
( , )
( , )
( , )
R i j
→
( , )
R i j
→
( , ),V i j
( , ),
( , ),
( , ),
( , ),
t i j
( , ),
V i j
( , )G i j
( , )
( , )
( , )
( , )
G i j
( , )
G i j
( , ),t i j
( , ),
( , ),
( , ),
( , ),
p i j
( , ),
p i j
=
( , )B i j
( , )
( , )
( , )
( , )
B i j
( , )
B i j
( , )p i j
( , )
( , )
( , )
( , )
V i j
( , )
q i j
=
i
i
i
i
H i j
H i j
H i j
H i j
p i j
q i j
V i j
V i j
V i j
H i j
S i j
f i j S i j
f i j S i j
−
→
⎡
⎣
⎤
⎥
⎦
=⎢
=−
=
=
=
−
−
−
=
=
=
=
=
=
=
=
=
=
=
=
==
=
i
i
i jq i j
p i j
p i j
G i j
G i j
G i j
V i j
V i j
q i j
=
=
B i j
B i j
B i j
p i j
t i j
V i j
H i j
H i j
R i j
R i j
=
→
→
==
=
=
(3)
Proc. of SPIE Vol. 6915 69153E-3

Page 4

where H(i, j) is the hue value and S(i, j) is the saturation value.

) , (
iB
),, (
i
), , (
iR
:
) , (
iB
,240
), (
MIN
), (
MAX
60
⎪
) , (
i
, 120
) , (
MIN
) , (
MAX
60
) , (
iR
, 0
) , (
MIN
) , (
MAX
60
jjGj MIN
MAX
MIN MAX
S
MAX
MIN MAX
S
jMAX if
jiGjiR
jGMAXif
jiRjiB
jMAX if
jiBj
iG
H
−
=
⎪
⎩
⎪
⎪
⎪⎪
⎨
⎪
⎪
⎧
−
=
=+
−
−
×
=+
−
−
×
=+
−
−
×
=

(4)
However, equation (2) cannot be used when S(i, j) is zero. The brightness was equally corrected from the center to the
skirt region of the fundus image (as shown in Fig.2 (b)). The color fundus images corrected by Bc(i, j) were then
processed by gamma correction. The gamma value was experimentally set to 1.5 (as shown in Fig.2 (c)). Finally, the
histograms of each red, blue, and blue-bit were extended (as shown in Fig.2 (d)). The fundus images were standardized
and made unclear by using these processes.


(a) (b)

(c) (d)

Fig. 2. Color contrast enhancement. (a) Original color fundus image. (b) Brightness of fundus image
was changed by the nonlinear curve with brightness values of HSV space. (c) Image after
processing by gamma correction. (d) Image was adjusted in the dynamic range.
Proc. of SPIE Vol. 6915 69153E-4

Page 5

—p
'.
/
I
I•I*r
I ç:c

2.4 Extraction of optic nerve head
The color images were converted into the grayscale images by selecting only the green component. Thus achieving a
clearer contrast was obtained as compared to the original images, which included all the color components. After the
optic nerve head was highlighted with brightness greater than that used for highlighting other tissues, it was investigated
by the p-tile method [10]. The shape of the detected region was approximated to a circle.

2.5 Detection of hemorrhage candidates
The pixel values of hemorrhages in an image are lower than those of other regions. Therefore, the haemorrhages
were detected by performing finite difference calculations along with smoothing. This method was carried out in two
steps: the rough and detailed detection processes. Firstly, the fundus images were smoothed by using a mask of 3 × 3
pixels. Next, the difference in the pixel values between two smoothed images was calculated. Subsequently, the
hemorrhage and blood vessel candidates were segmented by the thresholding technique. Fig. 3 (b) shows an image of
the roughly detected vessels. As shown in this image, the end of thin vessels was not detected. Therefore, the end of the
vessels was detected by using a similar method that uses two types of images smoothed with a mask of 9 × 9 pixels. Fig.
3 (c) shows an image of the detailed parts of the vessels. By combining both types of techniques for the detection of
hemorrhages and vessels, we could detect all the blood vessels from the optic nerve head up to the end of the vessels.
Fig. 3 (d) shows an image of the end of the vessels. Finally, the vessel candidates connected to the optic nerve head
were eliminated (as shown in Fig. 3 (e)) and the vessel candidates with large or very small areas were eliminated. Fig. 3
(f) shows an example of a resulting image.


(a) (b) (c)


(d) (e) (f)

Fig. 3. Illustration of the hemorrhage detection processes. (a) Input image. (b) Roughly detected
hemorrhages and blood vessels. (c) Details of detected hemorrhages and blood vessels. (d)
Combination of (b) and (c). (e) Candidates connecting blood vessels and candidate with large
areas were eliminated. (f) Final image of hemorrhage detection.
Proc. of SPIE Vol. 6915 69153E-5