SiiTAutomatic Exudates Detection on Thai Diabetic

Retinopathy Patients’ Retinal Images

Akara Sopharak1, Bunyarit Uyyanonvara2

1,2Sirindhorn International Institute of Technology, Thammasat University, THAILAND 1 akara@siit.tu.ac.th, 2 bunyarit@siit.tu.ac.th

ABSTRACT

Diabetic retinopathy is a complication of diabetes that is caused by changes in the blood vessels of the retina. The symptom can blur or distort the patient’s vision and are mainly cause of blindness. Exudates are one of the primary signs of diabetic retinopathy. Detection of exudates by ophthalmologists normally required pupil dilation using chemical solution which takes time and affects patients. This paper investigated a set of optimally adjusted morphological operators used for exudates detection on Thai patients’ non-dilated pupil and low-contrast images. The detected exudates are validated by comparing with ophthalmologists’ hand-drawn ground-truth. As a result, the sensitivity and specificity for the exudates detection were very high of 80% and 99.4% respectively.

Keywords: Diabetic retinopathy, Exudates, Retinal image

1. INTRODUCTION Diabetic-related eye diseases are the most common

cause of blindness in the world, especially on developing country such as Thailand. Patient’s sight can be affected by diabetes which causes cataracts, glaucoma, and most importantly, damage to blood vessels inside the eye, a condition known as "diabetic retinopathy". Diabetic retinopathy is a critical eye disease which can be regarded as manifestation of diabetes on the retina. It’s a major public health problem and it remains the leading cause of blindness in people of working age. The screening of diabetic patients for the development of diabetic retinopathy can potentially reduce the risk of blindness in these patients by 50% [1-4]. However, it is a normal situation in Thailand that the number of ophthalmologists is not sufficiently enough to cope with all patients, especially in Thai rural area [5]. So, automatic exudates detection would be necessarily useful in order to detect and treat diabetic retinopathy in an early stage.

G.G. Gardner et al. [6] proposed an automatic detection of diabetic retinopathy using an artificial neural network. The exudates are identified from grey level images. The fundus image was analyzed using a back propagation neural network. Zheng Liu, Chutatape Opas and Shankar M Krishnan [7] detected exudates using local thresholding. C. Sinthanayothin et al. [8] reported the result of an automated detection of diabetic

retinopathy on digital fundus image by recursive region growing segmentation (RRGS) algorithm. T. Walter et al. [9] detected exudates using grey level variation and means of morphological reconstruction techniques. Colour normalization and local contrast enhancement followed by fuzzy C-means clustering and neural networks were used by A. Osareh et al. [1].

Most of the techniques mentioned earlier worked on dilated pupils which the exudates and other retinal features are clearly visible. In normal diabetic retinopathy screening process, pupils were dilated using Tropicamide 1% eye drop. The process takes about 15-20 minutes to work and have an effect to patient. The examination time and patient’s effect could be reduced if the system can work on non-dilated pupil. Therefore, this paper proposed exudates detection on Thai retinal images and low quality images from non-dilated pupils.

2. METHODOLOGY We used digital retinal images obtained from

KOWA-7 non-mydriatic retina camera with a 45° field of view as our initial image dataset. The images were stored as jpeg image format (jpg) files with lowest compression rates. The image resolution is 500 x 752 pixels at 24 bit. The patient’s pupils were not dilated at the screening process.

2.1 Preprocessing Firstly, original image’s RGB space was transformed

to HSI space because HSI colour space is more appropriate since the intensity component is separated from the other two colour components. A median filtering operation was applied on the I band to reduce noise before a Contrast-Limited Adaptive Histogram Equalization [10] was applied for contrast enhancement. Exudates lesions and optic disc region are normally showing high intensity values in this channel and thus the contrast enhancement technique assigns them the highest intensity values [1, 2]. Original RGB image is shown in Fig.4 (a) and intensity channel after pre-processing is shown in Fig.4 (b).

2.2 Optic Disc Elimination Exudates detection is our main purpose, however we

have to remove the optic disc prior to the process because it is appearing in similar bright pattern, colour and

Proceedings of the 2006 ECTI International Conference, pp.709-712; May (2006)

SiiTcontrast [3, 9, 11]. Optic disc is characterized by the largest high contrast area. While vessels also appear in high contrast, the size of the area is much smaller. Applying closing operator (φ) on intensity channel (fI) will help eliminate the vessels which may remain in the optic disc region. Fig.4 (c) is a result after closing operator (Eq. 1) was applied.

( )

1 ( )sBIOP fφ= (1)

where sB is the morphological structuring element B of size s.

The resulting image was binarized by thresholding, as shown in Fig.4 (d). The thresholded image was then used as a mask. All the pixels in the mask were reversed before they were overlaid on the original image to remove candidate bright regions. The result, OP2, is shown in Fig.4 (e). The morphological reconstruction by dilation, R, was then applied on the previous overlaid image.

3 ( ) ( )If

OP x R OP= 2

)

(2)

The dilations of marker image (OP2) under mask image (fI) were repeated until the contour of marker image fits under the mask image. Reconstructed image is shown in Fig.4 (f). The difference between the original image and the reconstructed image was thresholded at grey level α1 as in Eq. (3). As a result, high intensities are reconstructed while the rest is removed, as shown in Fig.4 (g). Let T = {tmin,…,tmax} be an ordered set of grey levels, we have

1 max4 [ , ] 3(t IOP T f OPα= − (3)

The largest area, an optic disc, was identified and dilation operator (δ), Eq. (4), was applied in order to ensure that all optic disc areas are detected.

( )4(largest area ( ))sB

segOP OPδ= (4)

All optic disc regions were masked out using the previous output. The result is shown in Fig.4 (h).

2.3 Exudates Detection Proteins and lipids leaking from the blood into the

retina via damaged blood vessels is the main cause of exudates. They can be identified on the ophthalmoscope as areas with hard white or yellowish colours and varying sizes, shapes and locations, near the leaking capillaries within the retina [1].

Similar to the previous steps, high contrast vessels can be eliminated first by a closing operator before local variation operator, Eq. (5), was applied. E1 is an image after applying a closing operator. The result image, E2, is shown in Fig.5 (a).

1

22 1

( )

1( ) . ( ( ) ( ))1 E

i W xE x E i

Nμ

∈

= −− ∑ x (5)

where x is a set of all pixels in a sub-window W(x), N is a number of pixels in W(x), μE1(x) is the mean value of E1(i) and i∈W(x).

The resulting image was thresholded at grey level α2 in this step to get rid of all regions with low local variation. To ensure that all the neighbouring pixels were also included in the candidate region, dilation operator was also applied, as indicated in Eq. 6. The result is shown in Fig.5 (b).

[ ]( )2 max

( )3 2, ( )sB

tE T Eαδ= (6)

Exudates’ candidate regions should not be only their borders, the holes was flood-filled in this step, Fig.5 (c). The previously detected optic disc region was dilated before it was used to remove the optic disc from the above resulting flood-filled image (E4), Eq. 7. The result is shown in Fig.5 (d).

( )5 4 (sB )segE E OPδ= − (7)

The result from Eq.7 was used as a mask, showing all possible candidate regions of exudates, to create a marker image (E6), as shown in Fig.5 (e). It was later morphologically reconstructed using dilation operator under original intensity image. The result is displayed in Fig.5 (f).

Using Eq. 8, the final result is obtained by applying a threshold operation at grey level α3 to the difference between the original image (fI) and the reconstructed image (E7). The resulting image is shown in Fig.5 (g).

[ ]3 max 7, (seg ItE T f Eα )= − (8)

The result from this step will be sent for validation, as described in next section. Fig.5 (h) displays the result superimposed on original image. Other exudates detection examples are shown in Fig.6 and the overall procedures of exudates detection is shown in Fig.3.

3. EVALUATION Forty retinal images were processed by steps

proposed in the methodology section. The resulting detected exudates by the algorithm were compared with hand-drawn ground truth images created by expert ophthalmologists as shown in Fig.1. The sensitivity and specificity of the exudates detection, calculated based on its true positive, true negative, false positive and false negative are 80% and 99.4%, respectively. Sensitivity and specificity for all the test images are shown in Fig.2. Sensitivity are represented as circle points while specificity are represented as square points.

Proceedings of the 2006 ECTI International Conference, pp.709-712; May (2006)

SiiT

(a) (b)

Fig.1: Comparison of Exudates Detection (a) Ground Truth Image, (b) Detected Image.

0.00

0.20

0.40

0.60

0.80

1.00

1.20

1 4 7 10 13 16 19 22 25 28 31 34 37 40

%

Image

Number

Fig.2: Sensitivity and Specificity of Exudates Detection.

Create marker image

Clinical Validation

Result image =Original image - Reconstructed image

Threshold on Result image

Morphological Reconstruction

Remove Optic Disc

Fill the hole

Threshold image

Apply Local Variation

Closing

RGB to HSI

Apply median filter on I band

Apply contrast enhancement on I band

Optic Disc Elimination (2.2)

Retinal Image obtained from KOWA-7 non-mydriatic

retina camera

Exudates Detection

Fig.3: Procedure of Exudates Detection.

4. CONCLUSION In this work we have proposed a set of optimally

adjusted morphological steps to automatically detect optic disc and exudates from Thai diabetic retinopathy patient’s non-dilated pupil digital images. The results were then validated by comparing with ophthalmologists’ hand-drawn prediction. The outcome is quite successful with high sensitivity and specificity of 80% and 99.4%, respectively.

5. ACKNOWLEDGEMENT We would like to thank Eye Care Center,

Thammasart Hospital who supplied all the images used in this project.

6. REFERENCES [1] A. Osareh, M. Mirmehdi, B. Thomas and R.

Markham, “Automated Identification of Diabetic Retinal Exudates in Digital Colour Images,” British Journal of Ophthalmology, Vol.87, No.10, pp.1220-1223, 2003.

[2] C. Sinthanayothin, J.F. Boyce, H.L. Cook, T.H. Williamson, “Automated Localization of the Optic Disc, Fovea, and Retinal Blood Vessels from Digital Colour Fundus Images,” Br. J. Ophthalmal, Vol.83, pp.231-238, 1999.

[3] C.I. Sanchez, R. Hornero, M.I. Lopez, J. Poza, “Retinal Image Analysis to Detect and Quantify Lesions Associated with Diabetic Retinopathy,” Proceedings of 26th IEEE Annual International Conference on Engineering in Medicine and Biology Society (EMBC), Vol.1, pp.1624 – 1627, 2004.

[4] W. Hsu, P.M.D.S Pallawala, Mong Li Lee, Kah-Guan Au Eong, “The Role of Domain Knowledge in the Detection of Retinal Hard Exudates,” Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, Vol.2, pp.II-246 - II-251, 2001.

[5] R. Paisan, W. Nattapon, S. Pattanaporn, P. Ekchai, T. Montip, “Screning for Diabetic Retinopathy in Rural Area Using Single-Field, Digital Fundus Images,” J. Med Assoc Thai, Vol.88, pp.176-180, 2005.

[6] G.G. Gardner, D. Keating, T.H. Williamson, A.T. Elliot, “Automatic Detection of Diabetic Retinopathy using an Artificial Neural Network: a Screening Tool,” Br. J. Ophthalmol, Vol.80, pp.940-944, 1996.

[7] Zheng Liu, Chutatape Opas, Shankar M Krishna, “Automatic Image Analysis of Fundus Photograph,” Proceedings of 19th IEEE International Conference on Engineering in Medicine and Biology (EMBS'97), Vol.2, pp.524-525, 1997.

[8] C. Sinthanayothin, J.F. Boyce, T.H. Williamson, H.L. Cook, “Automated Detection of Diabetic Retinopathy on Digital Fundus Image,” International Journal of Diabetic Medicine, Vol.19, pp.105-112, 2002.

[9] T. Walter, J.C. Klein, P. Massin, A. Erginay, “A Contribution of Image Processing to the Diagnosis of Diabetic Retinopathy-Detection of Exudates in Colour Fundus Images of the Human Retina,” IEEE Transactions on Medical Imaging, Vol.21, pp.1236 – 1243, 2002.

[10] R.C. Gonzales, R.E. Woods, Digital Image Processing, Addison-Wesley publishing Co., 1993, ch.3,9,10.

[11] Zhang Xiaohui, Opas Chutatape, “Detection and Classification of Bright Lesions in Colour Fundus Images,” Proceeding of IEEE International conference on Image Processing (ICIP), Vol.1, pp.139-142, 2004.

Proceedings of the 2006 ECTI International Conference, pp.709-712; May (2006)

SiiT

(a) (b) (c) (d)

(e) (f) (g) (h)

Fig.4: Optic Disc Location (a) RGB Image, (b) I band after Pre-processing, (c) Image after Closing, (d) Thresholded Image, (e)Marker Image, (f) Reconstructed Image, (g) Threshold on Difference Image, (h) Optic Disc Area.

(a) (b) (c) (d)

(e) (f) (g) (h)

Fig.5: Exudates Detection (a) Applied Local Variation, ( b) Thresholded Image, (c) Filled Image, (d) Remove Optic Disc from Image, (e) Marker Image, (f) Reconstructed Image, (g) Difference Image, (h) Superimposed on Original.

(a) (b) (c)

(d) (e) (f)

Fig.6: Exudates Detection on Low Quality Images (a), (d) Original Images, (b), (e) Detected Exudates, (c), (f) Detected Exudates Superimposed on Original Images.

Proceedings of the 2006 ECTI International Conference, pp.709-712; May (2006)