[IEEE 2013 IEEE International Conference on Control System, Computing and Engineering (ICCSCE) - Penang, Malaysia (2013.11.29-2013.12.1)] 2013 IEEE International Conference on Control

Non-Invasive Contrast Enhancement for Retinal Fundus Imaging

Ahmad Fadzil M Hani, Toufique.A.Soomro, Ibrahima Fayee Centre for Intelligent Signal and Imaging Research,

Universiti Teknologi PETRONAS, 31750 Tronoh, Malaysia

Abstract— Diabetic Retinopathy (DR) is vision loss impairment due to complication arising from diabetic condition affecting the retina. It is known that the foveal avascular zone (FAZ) enlarges with progression of DR due to the loss of capillaries in perifoveal capillary network. However, normal retinal fundus images suffer from low and varied contrast problems. A non-invasive digital image enhancement technique called RETICA has been developed that overcomes the problem of varied and low contrast in fundus images. RETICA first normalises the varied contrast using a Retinex-based method that separates the illumination from the reflectance part of the image followed by ICA that forms the original retinal pigment makeup namely the macular, haemoglobin and melanin retinal pigment. The technique has been applied on our FINDeRS dataset contained 175 fundus images and another 35 fundus image pairs obtained from an earlier study containing colour fundus images and their corresponding fluorescein fundus angiogram (FFA) images. For the 35 image pairs, RETICA achieved an average contrast improvement factor (CIF) of up to 5.46 compared to the invasive FFA at 5.12. For the FINDeRS images, RETICA achieved an average CIF of 5.63 with denoising. The RETICA image enhancement technique potentially reduces the need for the invasive fluorescein angiogram in DR assessment.

Keywords—Contrast Enhancement; diabetic retinopathy; Retinex; Independent Component Analysis; RETICA

I. INTRODUCTION

Eye screening using fundus imaging is necessary and important for detecting and monitoring of diabetic retinopathy (DR). The main purpose of the screening is to recognise subjects with sight-threatening DR so that essential treatment would be given for prevention of vision loss. The progression of DR starts from mild to severe and ends with a complete vision loss [1]. Retinopathy has two categories generally, non-proliferative diabetic retinopathy (NPDR) and proliferative diabetic retinopathy (PDR). These DR categories are characterised by the presence of pathologies such as haemorrhages, exudates and changes in the veins. NPDR contains mild, moderate and severe NPDR [2]. It has been observed that retinal vasculature features analysed in DR are the loss of retinal capillaries in the perifoveal capillary network due to the increase in the size of the Foveal avascular zone (FAZ) [3]. Recent research on the analysis of fundus images found that the size of the fovea avascular zone increased with the severity level of DR. The FAZ can be observed in colour fundus images

and the perifoveal capillary network can only be seen in fundus fluorescein angiograms (FFA) [4].

The goal of the present study is to address the issues of the low and varied contrast image when retinal imaging is performed by colour fundus camera but without contrasting agents injected into subject. Moreover, the current protocol of evaluating DR is based on the FFA images [5]. A digital image enhancement called RETICA, to extract the retinal vasculature from colour fundus images has been developed [6]. The method avoids the need of fundus fluorescein angiograms (FFA) [7]. In this work, the contrast improvement factor of FINDeRS database (175 colour fundus images) and 35 image pairs of FFA and digital colour fundus images have been further investigated using RETICA.

II. PROPOSED APPROACH

The proposed digital image enhancement technique is the combination of the two techniques Retinex and Independent Component Analysis (ICA), also known as RETICA [5]. The first stage of RETICA involves contrast normalization through the Retinex algorithm [8] and second stage performs contrast enhancement of object of interest using Independent Component Analysis (ICA) [8] method as depicted in Figure 1. The Red, Green, and Blue colour channels of the fundus image are processed by the Retinex algorithm to normalise the varied contrast. Next, ICA is performed to obtain the independent components (due to macular, haemoglobin and melanin pigments) from the colour channels. The haemoglobin image is selected from the three independent components. Details of the processes are explained in the following subsections.

Contrast Normalisation: As shown in Figure 1, the RGB colour channels of the fundus image are each processed through Retinex algorithm to get the corresponding normalised RGB colour channels. A smaller region of the fundus image containing the macular area is used to investigate the enhancement of the retinal capillaries.

Description of Retinex Algorithm: McCann et al [9] improved the random walk Retinex algorithm by developing the multilevel one dimensional Retinex algorithm [10]. McCann introduced the iterative Retinex concept based on multi-resolution pyramid concepts and iteration process

2013 IEEE International Conference on Control System, Computing and Engineering, 29 Nov. - 1 Dec. 2013, Penang, Malaysia

978-1-4799-1508-8/13/$31.00 ©2013 IEEE 197

Macular Region

Fundus Image

RGB of maculaRegion

R

through ratio-product-reset-operation. The ala multi-resolution pyramid from the input bimage data forming pyramid layers of datFigure 2.

A fundamental concept behind the Retinex lightness is the comparison of pixel value pixels of image. In multi-resolution pcomparison starts at the most averaged top leAfter computing lightness of the imagresolution (top level pyramid), the resulting are propagated down, by pixel replication tnext level as initial lightness estimates at pixel comparison process continues to refinestimates down a next level as initial lightnethe new product or refine estimated lightnebottom level of pyramid ( original image reso

The iterative Retinex algorithm procesaccording to multi-resolution pyramid bdepend upon the number of iteration. Retinex calculates the long distance iterationmoves to short distance iterations. At each sbetween the pixels being compared decreasepixel distance. The direction among the pixeach step in clockwise direction. At each scomparison is implemented to estimate rusing ratio-product-reset–average operationiteratively computed with certain number number of iterations is a very important pRetinex algorithm.

Implementation of McCann99 Retinex Aimplementation of McCann99 contains four flow chart of McCann99 algorithm is shown

1. The input images must be of dim. 2 where w≥h and w and h arerange. This limitation arises from thlevel of the image pyramid differslevels by a factor of 2 in each dimein Figure 2.

ar Retinex Algorithm ICA

R B

G

lgorithm creates by averaging the ta as shown in

computation of to that of other pyramid, pixel evel of pyramid. ge at reduced lightness values o the pyramid’s that level. The

ne the lightness ss estimate until ess computed at olution) [8].

ssed image data ut the process These iterative

ns then gradually step, the spacing es with one-shift els also alters at step, the `pixels reflectance part

n [11], which is of times. The

arameter of the

Algorithm: The below steps and in Figure 3 [8].

mension . 2e integers in the he fact that each s from previous ension as shown

2. In

the first step, the log

lowest resolution lelevel will be doublepyramid depends oThe number of laye2, dividing both theimages as calculatedin the Matlab.

3. When the results (caof dimension n×m hare then replicated tdimension 2n ×2m.

4. At all levels of pcalculated for calcueach pixel is compuimmediately neighorder. Each visit average operation, function CompareWsubtracts the neighbstep) and then add(the product step orexceeds the maximreset to Maximum product for the pixeneighbour is averproduct (average soperation is shown i

Figure 1 RETICA Method

Figure 2 Im

Haemoglobin Image

RETICA Image

g image is averaged down to the

vel. At each step, the resolution ed. The number of layers in the on the size of the input image. ers will be the greatest power of e width and height of the input d by the function ComputeSteps

alled new products) at one level have been computed, the values to form an old product image of

pyramid, the new product are ulated estimated lightness and

uted by visiting each of its eight hbouring pixels in clockwise involves a ratio-product-reset-which is implemented by the

WithNeighbor in Matlab. It bour’s log luminance (the ratio s the result to the old product

r shift input image). If the result mum defined by Maximum, it is

(the reset step) finally, the new el obtained by comparison to its raged with the previous old steps) and mathematically this in Equation 1 and in Figure 4.

mage Pyramid


198

A fundamental parameter to the McCann99 algorithm is the number of times a pixel’s neighbours are to be visited. This is done by iteration. Iteration controls the number of times the neighbours are cycled through, which, as a result, affects the distance at which pixels influence one another. This happens because the new product values for all pixels are being computed in parallel, for second iteration, all neighbouring pixels have had their new products update values. In second iteration, these update values or new values contain information propagated from beyond pixels immediate neighbours. This process is continue with given number of iterations output image is obtained that is simply the input image scaled by the image maximum value. Retinex Algorithm for Retinal Fundus Image: The McCann iterative Retinex algorithm [8] chosen for contrast normalisation of fundus image. Figure 3 shows the flow chart of McCann99 iterative Retinex [6]. RGB input image is first decomposed into the three channels and each channel is processed by the Retinex algorithm. Each colour channel input image is transformed from linear to logarithmic form to simplify the process; multiplication to addition and division to subtraction. The multi-resolution pyramid of each colour channel is created.

In each step, the spacing between the pixels being compared decreases with one-shift pixel distance. The direction among the pixels also alters clockwise at each step. At each step, the pixels comparison is implemented to estimate reflectance part using ratio-product-reset–average operation [11]. Subsequent comparison between pixels are continuously performed to refine the estimated reflectance at pyramid’s bottom level until the spacing decrease to one pixels and final product is achieved.

Ratio and product are process of accumulating and comparing resulting in the revision of a newer product or initial estimated product in each process of pixel comparison. Reset operation is to normalise the newer product exceeding the sustainable maximum (one pixel shift distance is maximum sustainable). Averaging process aims to estimate and update one pixel’s luminance.

The ratio-product-reset-average operation is performed by calculating the ratio between I (in specific channel) and its spatially shifted input version and offset by some distances formulated as Equation 1. logR , R I , I , R , R ,

Where logI , logI , represented the ratio and logI , logI , logR , represents the product in log domain. Reset operation is performed to update the maximum intensity according to number iteration. The logR , is a result of averaging with logR , and logR , itself is an updated output produce in each iteration that will be used as an input for next iteration till the final reflectance is obtained at given last iteration.

Figure 3 McCann99 Algorithm Flow chart

Figure 4 Flow chart of Ration-Product-Reset-Average Operation

(1)


199

Contrast Enhancement: The FastICA algsymmetrical orthogonalisation is used to geindependent components because of its goohigh computational speed for high dimensThe output image of the Retinex algorithImages) is fed as inputs to the ICA as showThe ICA is a technique to determine the from the mixtures of several independent scase, the enhancement of the low contrasblood vessels in the digital fundus image idetermining the retinal pigment makhaemoglobinIH, melanin IM and macular IMCthe ICA. The independent component due exhibits higher contrast of retinal bloobackground (melanin and macular componen

Evaluation of RETICA Performance: Tare used to evaluate the RETICA performameasurement of the contrast between the blbackground of macular region of green baimage, Haemoglobin image. Second isimprovement factor of RETICA versus greand FFA versus green band image.

Contrast determination: The contrast of tvessels and the background region is the intensity difference between the retinal blothe background of the retinal image andaccording to Equation 2.

C| | 1n I I

Here, C| | is the contrast between thvessels and background. The terms I and intensities of the retinal blood vessels and respectively. The n variable indicates the nuof the retinal blood vessels and backgrounimage. In this study, n= 50 and the pixeselected randomly.

Selection of pixels for data intensity: Tcontrast of the image, the intensity points oblood vessels and background within the mselected randomly as depicted by the blue dSimilarly, the intensities of the retinal bloodbackground of macular region of green baimage and haemoglobin image due to RETIrandomly. Contrast between the bloobackground of macular region of green bandregion of grey scale of FFA image and mahaemoglobin .image is measured according t

Selection of Haemoglobin Image: The simage due to the haemoglobin is performeby comparing the kurtosis of independent com

gorithm with the et the estimated od accuracy and ional data [12].

hm (Normalised wn in Figure 1. original signals

sources. In this st of the retinal is performed by ke-up, namely C pigment using to haemoglobin

od vessels and nts).

Two parameters nce. First is the ood vessels and

and image, FFA s the contrast een band image

the retinal blood absolute mean

ood vessels and d is determined

he retinal blood I refer to the its background, umber of pixels d in the fundus

els locations are

To measure the or pixels of the

macular region is dots in Figure 5. d vessels and the and image, FFA CA are selected

od vessels and d image, macular acular region of to Equation 1.

selection of the ed automatically mponent image.

Kurtosis is used to differentifrom other two components (Pigments). Kurtosis is medistribution of data and lReferring to Figure 1, thaemoglobin related IC as thkurtosis [13]. Because ofhaemoglobin image (Retinal background (either melanbackground) of retinal fuhaemoglobin image should bICs (melanin and macular pig

Contrast Improvement Facfactor (CIF) of the FFA imathe contrast between the grereference green band imimprovement factor (CIF) RETICA image is definedbetween the haemoglobin iband image as formulated Ta

Table 1Contrast I

CIFRETICA represents conRETICA and CIFFFA repres

Figure 5 Data intensity of bselectio

Green Band Image

Green Band Blood vintensity selectio

FFA Image

FFA Blood vessels inselection

RETICA Image

Haemoglobin Blood intensity Selectio

RETICACIFRETICA CRETICCREF

(2)

ate the haemoglobin related ICs (Macular Pigments and Melanin easured of how peaked or flat leads to haemoglobin image. the algorithm determines the he IC with the least value of the f the smaller area (pixels) of blood vessels) compared to the nin background or macular

undus image, the kurtosis of be smaller than that of other two gment).

ctor: The contrast improvement age is calculated as the ratio of ey scales of the image and the mage. Whilst, the contrast

of the haemoglobin of the d as the ratio of the contrast image and the reference green able 1.

Improvement Factor

ntrast improvement factor of ents the contrast improvement

blood vessels and Background on Method

essels on

Green Background Intensity Selection

ntensity FFA Background Intensity Selection

vessel on

Haemoglobin Background Intensity selection

FFAA CIFFFA CFFACREF


200

factor of FFA. Here,CRETICA is the contrast of RETICA image and CFFA is the contrast of the FFA image. The CREF is the contrast of the reference image which is green band image.

III. RESULTS AND DISCUSSION

In this work, RETICA is an applied on our FINDeRS (Fundus Image for Non-invasive Diabetic Retinopathy) database that contained 175 images. RETICA is an applied another database that contained 35 image pairs (colour fundus image and FFA image). Table 2 shows the contrast values of green band (CGB), FFA (CFFA), and RETICA or haemoglobin (CHI) images, and the CIF achieved for FFA (CIFFFA) and haemoglobin (CIFHI) of the 35 image pairs.

Table 2 Contrast and CIF of 35 Retinal Image Pairs

No Image CGB CFFA CHI CIFFFA CIFHI

1 No DR_1 6 33.8 34.4 5.63 5.73 2 No DR_2 6.9 38.3 39.4 5.55 5.71 3 No DR_3 5.6 28.4 31.1 5.07 5.55 4 No DR_4 6.3 31.6 35.8 5.01 5.68 5 No DR_5 6.2 33.2 35.5 5.35 5.72

6 No DR_6 11.7 53.3 60.1 4.55 5.13

7 No DR_7 6.4 34.9 35.3 5.45 5.51 8 No DR_8 9.1 47.7 48.2 5.24 5.29 9 No DR_9 8.5 34.3 47.5 4.03 5.58

10 No DR_10 7.1 40.9 36.9 5.76 5.19 11 No DR_11 8.7 35 45.5 4.02 5.22

12 Mild_1 5.2 28.7 29.9 5.51 5.75

13 Mild_2 5.3 29.2 26.9 5.50 5.07 14 Mild_3 12.9 73 68.5 5.65 5.31 15 Mild_4 12.9 50.2 67.1 3.89 5.20 16 Mild_5 5.1 26.1 29.3 5.11 5.74 17 Mild_6 10.7 54.9 58.9 5.13 5.50

18 Moderate_1 5.6 28.3 30 5.05 5.35

19 Moderate_2 5.1 27.6 26.5 5.41 5.19 20 Moderate_3 7.8 44.4 39.2 5.69 5.02 21 Moderate_4 9.3 49.5 52.8 5.32 5.67 22 Moderate_5 14.2 75.5 82.2 5.31 5.78 23 Moderate_6 9.9 42.1 53.8 4.25 5.43

24 Severe_1 6.5 34.9 36.1 5.36 5.55

25 Severe_2 5.1 29.2 31.2 5.72 6.11 26 Severe_3 10.9 54.2 54.9 4.97 5.03 27 Severe_4 11.2 59.8 59.9 5.33 5.34 28 PDR_1 5.2 26.5 29.8 5.09 5.73 29 PDR_2 5.2 27.1 29.6 5.20 5.69

30 PDR_3 3.9 14.9 21.1 3.82 5.4

31 PDR_4 6.3 32 31.8 5.07 5.04 32 PDR_5 9.5 42.8 48.7 4.50 5.12 33 PDR_6 4.6 27.2 25.6 5.91 5.56

34 PDR_7 15.8 87.3 82.2 5.52 5.20 35 PDR_8 7.4 38.5 43.4 5.20 5.86

Average 7.9 40.4 43.1 5.12 5.46

As shown in Table 2, it is observed that RETICA achieved higher average contrast 43.1 as compared to FFA method 40.4. RETICA achieved average CIF of 5.46 as compared to 5.12 of FFA. Generally, this improvement in the contrast in colour retinal fundus image through non-invasive digital image enhancement method potentially reduces the need for the invasive fluorescein angiogram in DR assessment.

Referring Table 2, it is observed that some images pairs have very high contrast values (contrast is above 50 of RETICA and FFA, and above 10 of green band images) as highlighted in the table.

Referring to Figure 6, the green band of No_DR_6 image shows clearly blood vessels with contrast of 11.17. The RETICA image is also more enhanced image with retinal blood vessels more observable as compared to FFA image; RETICA image have the contrast of 60.1 as compared to FFA 53.3. Similarly, RETICA images of Mild_4, Mild_6, Moderate_5, Severe_3 and Severe_4 clearly show tiny capillaries as compared to their corresponding FFA images. For Mild_3 and PDR_7, the FFA gave better contrast compared to RETICA. It can be seen that the retinal blood vessels are so much brighter due injecting contrast agent however RETICA image still gave equally good observation of blood vessels.

Referring Table 3, the contrast improvement factor of 35 images pairs at each DR-severity is shown for all DR severity stages; RETICA outperformed the FFA in improving the contrast of the green band images as shown in Figure 7.

Table 3 CIF of FFA and RETICA for DR severity levels

DR-Stage FFA RETICA No_DR 5.06 5.49

Mild NPDR 5.13 5.43 Moderate NPDR 5.17 5.41

Severe NPDR 5.35 5.51 PDR 5.04 5.45

Average CIF 5.12 5.46

Figure 8 shows the performance of RETICA on the FINDeRs database. It is seen that the contrast improvement drops with the DR severity. For FINDeRS database, RETICA gave an average contrast improvement factor of 4.98. The main reason of drop contrast improvement factor is due to noise in the Retinex image. After denoising using stationary wavelet transform technique on Retinex image, contrast improvement factor above 5 in all DR stages.


201

Image Green Band FFA Image

No_DR_6

Contrast: 11.17 Contrast: 53.3

Mild_3

Contrast: 12.9 Contrast:73

Mild_4


Mild_6


Moderate_5


Severe_3


Severe_4


PDR_7


Figure 6 Comparison of selected FFA and R

Figure 7 CIF of RETICA and FFA

Figure 8. Average CIF of FINDeRS Database befRetinex Image

VI. CON

Analysing retinal fundustheir very low and varied cocontrast between the bloodmake it difficult to deteraccurately. Fluorescein angioimages, it is not preferablewhich injecting a contrast agthe validation of developedcalled RETICA on real capabilities of Retinex and able to normalize the varied low contrast of the fundus icontrast achieved by the REto reduce the use of invasivFor the 36 image pairs, Rcontrast improvement factorinvasive FFA at 5.15. Fodenoised Retinex image, it improvement factor of 5.6technique potentially reducfluorescein angiogram in DR

REFE[1] X. Zhitao, W. Jun, Z. Qia

Fundus Image Processing Automation and Systems Conference on, 2011, pp. 1-

[2] M. Niemeijer, B. van GinneI. Sanchez, et al., "Retinopof Microaneurysms in DigImaging, IEEE Transactions

[3] P. P. Goh, "Status of Diabeto the Diabetic Eye RegistryMalaysia, vol. vol. 63, pp. 2

[4] H. A. Nugroho, "Non-InvaFundus Image for ComputeGrading System," Phd TheProgramme , Universiti Tek

[5] P. D. Malaysia, "Clinical Type 2 Diabetes Mellitus,Endocrine and Metabolic So

[6] A. F. M. Hani, T. Ahmed "Enhancement of colour funin Biomedical Engineering Conference on, 2012, pp. 83

[7] M. Ahmad Fadzil, L. Izhar,retinal fundus images for Medical and Biological Eng700, 2011.

[8] B. Funt, F. Ciurea, and J. MElectronic Imaging, vol. 13,

[9] E. H. Land and J. J. McCanof The Optical Society of Am

[10] J. Frankle and J. McCann, M1983.

[11] R. Sobol, "Improving the Rrange photographs," Journa2004.

[12] A. a. O. Hyvarinen, E., , "Iand applications,," Neural N

[13] J. F. a. S. Cardoso, A., , "BIEEE proceedings-F,, vol. 1

Haemoglobin Image

Contrast:60.1

Contrast: 68.5

Contrast: 58.9

Contrast: 58.9

Contrast: 82.2

Contrast: 54.9

Contrast: 59.9

Contrast: 82.2

RETICA images

A

fore and After Denoised

NCLUSIONS

s images is difficult because of ontrast characteristics. The low d vessels and the background rmine the retinal vasculature ography produces high contrast

e due to its invasive nature in gent is a necessity. In this work, d digital enhancement method

fundus images exploits the ICA. Retinex based method is contrast as well as enhance the

image. The improvement of the TICA is significantly important

ve procedures such as the FFA. RETICA achieved an average r of up to 5.45 compared to the or the FINDeRS images after

achieved an average contrast 63. This image enhancement ces the need for the invasive R assessment.

RENCES an, and X. Jiangtao, "Diabetic Retinopathy Based on Phase Information," in Control, Engineering (CASE), 2011 International

3. eken, M. J. Cree, A. Mizutani, G. Quellec, C. athy Online Challenge: Automatic Detection gital Color Fundus Photographs," Medical s on, vol. 29, pp. 185-195, 2010. etic Retinopathy Among Diabetics Registered y,National Eye Database," The Med Journal of 24-28 , 200, 2007. asive Image Enhancement of Colour Retinal erised Diabetic Retinopathy Monitoring and esis Electrical and Electronics Engineering

knologi PETRONAS, 2012. Practice Guidelines (CPG) Management of

," Ministry of Health Malaysia, Malaysian ociety, Academy of Medicine Malaysia, 2009.

Soomro, H. Nugroho, and H. A. Nugroho, ndus image and FFA image using RETICA," and Sciences (IECBES), 2012 IEEE EMBS

31-836. , H. Nugroho, and H. Nugroho, "Analysis of

grading of diabetic retinopathy severity," gineering and Computing, vol. 49, pp. 693-

McCann, "Retinex in MATLAB™," Journal of pp. 48-57, 2004. nn, "Lightness and Retinex Theory," Journal

merica, vol. 61, 1971. Method and apparatus for lightness imaging,

Retinex algorithm for rendering wide dynamic al of Electronic Imaging, vol. 13, pp. 65-74,

Independent Component analysis: alogrithms Networks, vol. 13, 2000.

lind beamforming for non-Gaussian signals," 40, 1999.


202

Documents

[IEEE 2013 IEEE International Conference on Control System, Computing and Engineering (ICCSCE) - Penang, Malaysia (2013.11.29-2013.12.1)] 2013 IEEE International Conference on Control