Principal Component Analysis based Colour Scheme Optimisation in

Eye Fundus Images

Contrast Enhancement for Detection and Evaluation of Drusen in Age Related

Macular Degeneration Patients' Follow up

Algimantas Kriščiukaitis

, Robertas Petrolis

, Daiva Stanislovaitienė

and Dalia Žaliunienė

Lithuanian University of Health Sciences, Eiveniu str. 4, Kaunas, Lithuania

Eye Clinics of Lithuanian University of Health Sciences, Eiveniu str. 2, Kaunas, Lithuania

Keywords: Principal Component Analysis, Eye Fundus Imaging, Drusen, Age Related Macular Degeneration.

Abstract: Efficiency of the patient status monitoring in Age Related Macular Degeneration cases, based on evaluation

of morphological properties eye fundus images, can be significantly increased by specific contrast

enhancement in the images. Objects of interest - drusen (focal deposits of extracellular debris located

between the basal lamina of the retinal pigment epithelium and the inner collagenous layer of Bruch

membrane) usually are represented by various intensity but the same unique color in the image.

Construction of the optimal color scheme to increase the contrast of drusen can be realized by means of

Principal Component Analysis, which transforms original RGB color representation into principal

components space. The study demonstrates that proposed method can increase contrast-to-noise ratio of the

drusen areas 10-fold or more.

1 INTRODUCTION

Age related macular degeneration (ARMD) is a

degenerative disease usually occurring in people

over the age of 50 years. Condition and severity of

ARMD is classified according to diagnostic features

obtained from eye-fundus images. Such features are

estimates characterizing morphology and area

covered by drusen, focal deposits of extracellular

debris located between the basal lamina of the

retinal pigment epithelium (RPE) and the inner

collagenous layer of Bruch membrane (Spaide,

2010). The main forms of ARMD are defined

according to specific lesions, the characteristics of

drusen plays an important role here (Bird, 1995).

Evaluation of drusen properties is used for

monitoring of patient status dynamics. Area covered

by drusen and sizes of them are usually estimated by

means of various heuristic morphometric algorithms.

Efficiency of such algorithms mostly depends on

contrast between drusen area and background.

Therefore preprocessing of images usually starts

from selection of optimal spectral domain

maximizing this contrast. Spaide with coauthors

(Spaide, 2010) introduces optico-physical model

based spectral characteristics for differentiation of

drusen types classified by Gass (Gass, 1997):

discrete yellow-white punctate elevations (“hard”

drusen); large pale-yellow “placoid or dome-shaped

structures” (“soft” drusen) are seen in the eye fundus

singly or in groups. The color of the drusen depends

on its main substance and optical characteristics of

the covering layers. It is reported that all types of

drusen contain one main substance called

“lipoprotein-derived debris”, a lipid-rich material

(Curcio, 2009); (Russel, 2000). So it is expected that

the certain type of drusen observed in particular eye

fundus image probably will have unique, but the

same color. Due to its structure and location in

regard to illuminating light source and camera the

intensity could vary, however the color will remain

the same.

Determination of optimized color combinations

for blood vessels detection in eye fundus images is

reported in (Patasius, 2009). The performance of the

method was estimated by sensitivity (proportion of

correctly identified blood vessel pixels) and

specificity (proportion of correctly identified non-

blood vessel pixels). Optimization algorithm

maximizing area under ROC curve obtained using

Kriš

ciukaitis A., Petrolis R., Stanislovaitien

e D. and Žaliunien

e D..

Principal Component Analysis based Colour Scheme Optimisation in Eye Fundus Images - Contrast Enhancement for Detection and Evaluation of

Drusen in Age Related Macular Degeneration Patients’ Follow up.

DOI: 10.5220/0004910500660069

In Proceedings of the International Conference on Bioimaging (BIOIMAGING-2014), pages 66-69

ISBN: 978-989-758-014-7

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

training set images gives optimized color scheme.

Determination of special color scheme for detection

of drusen described in (Patasius, 2009-2). However

efficiency of universal predetermined color scheme

is sensitive to particular image registering conditions

and color distortions (e.g. white balance).

Individualized color scheme for every particular eye

fundus image can reveal even more important

diagnostic details related with preliminary detected

drusen. In this study we aimed to elaborate such

method.

The ordinary pixel of the color image is

represented as point in orthogonal RGB space. All

pixels of the same color but different intensity will

appear along certain direction in the RGB. It is

possible to find an optimal transform of given RGB

space that one of the axes of new coordinate system

will go along this direction representing color of the

drusen. Such transform could be found using

Principal Component Analysis (PCA). The aim of

this study was to show how PCA could be used to

find an optimal transform of representation of eye

fundus image in RGB space into optimal space,

maximizing contrast of drusen.

2 METHODS

Eye fundus images were taken using fundus camera

(Carl Zeiss Meditec AG, Germany) in 134 ARMD

patients who underwent treatment in Eye Clinics of

Lithuanian University of Health Sciences. Image

processing was performed using elaborated

programs in MatLab computation environment.

The experts (experienced Ophthalmologists)

have marked drusen areas in original eye fundus

images for further analysis. Contrast-to-noise ratio

of representation of drusen was evaluated according

following criteria:

C 

 M

(1)

where M

- mean of pixel values in drusen area; M

mean of pixel values in non-drusen area; s

standard deviation of pixel values of non-drusen

area.

Original representation of eye fundus images -

three-dimensional arrays representing pixel values in

red, green and blue colors, were transformed into

two dimensional arrays, concatenating all rows of

one color of the image into one. The resulting array

consisted of three rows representing pixel values in

red, green and blue colors respectively:

X 

r 2

... x













(2)

Principal Component Analysis transforms original

representation into new space of variables

maximizing variation and concentrating correlated

original variables (Jollife, 2002). If pixel values of

drusen area make enough big contribution to total

variance of image pixel values, we can expect that

first, or at least one of the first new variables

(principal components) will maximize contrast of

drusen areas versus the rest of the image. Spatial

correlation of original image representation X can be

estimated as:



X  X

(3)

The eigenvector equation for R

is:











(4)

where Λ denotes the eigenvalue matrix with the

eigenvalues sorted in descending order, and Ψ is the

corresponding eigenvector matrix. The matrix Ψ

defines an orthonormal transformation, which is

applied to the original data X







(5)

to obtain the transformed representation, rows of

which contain principal components of X.

Principal components were trasformed back to

two-dimensional arrays and shown to the experts.

Contrast-to-noise ratio was evaluated according to

formula (1) in principal component in which drusen

areas were most clearly visible.

3 RESULTS

Example of typical eye fundus image containing

drusen presented in Figure 1 and original

representation of its pixel values in orthogonal RGB

space on Figure 2. Pixel values of drusen areas form

prolonged cluster marked by the arrow. As one can

see, maximal variance of this part of the pixels is in

the direction close to the direction of G axis. It

complies with the results reported by (Patasius,

2009-2) that biggest part of information for drusen

detection should be taken from green color. On the

other hand, it confirms our expectations that drusen

areas due to their physicochemical properties

PrincipalComponentAnalysisbasedColourSchemeOptimisationinEyeFundusImages-ContrastEnhancementfor

DetectionandEvaluationofDruseninAgeRelatedMacularDegenerationPatients'Followup

probably will have different intensity but the same

color. So we can expect that PCA will find an

optimal representation space where one axis will go

along this linearly looking cluster of the pixel

values.

Figure 1: Original representation of the eye fundus image.

Drusen are white spots on the image marked by arrow.

Figure 2: Original representation of the pixel values of eye

fundus image in RGB space. Drusen area pixels form

linearly prolonged cluster.

Figure 3 illustrates typical result of PCA on eye

fundus image containing drusen. Three images in the

left column represent red, green and blue

components of original image. Right column of the

images represents three principal components of this

image. As one can see, maximal visual difference in

intensity between drusen area and the rest of the

image is in the first principal component. It looks

much bigger then in the green component of the

original image, recommended by other authors.

Eigenvector matrix calculated for the image

shown in figure 1 is presented in table 1. Values

used to construct first principal component of this

image (shown in italics) correspond to the

coefficients for construction of optimal color

combination proposed by (Patasius, 2009-2): 0.0287;

0.6975; -0.2738 for R, G and B components

respectively. Eigenvector values of the other

analysed pictures were in the same range.

Figure 3: Example of PCA on typical eye fundus image

containing drusen. Original R, G and B components of the

image presented on the left column. Three principal

components of the image are prsented on the right column.

Table 1: Eigenvector matrix calculated for image

presented in Figure 1.

-0,15

0,62 0,77

0,68

-0,5 0,53

-0,72

-0,61 0,35

Values of the contrast-to-noise estimates for this

image are presented in table 2. The maximal

contrast-to-noise estimate value 4.37 is for green

component of the original image, however the same

estimate reaches 98.56 for Y1 principal component,

showing significant increase in contrast after PCA

procedure. In majority of images increase in contrast

was at least 10 fold or more.

Table 2: Mean pixel values of drusen and non-drusen

areas and contrast-to-noise estimates C.

Drusen Non-drusen C

Mean StDev Mean StDev

R 237,84 84,89 173,52 581,13 2,67

G 184,81 796 116,07 247,82 4,37

B 109 505,44 74,68 92,96 3,56

Y1 248,1 137,29 39,84 4,46

98,56

Y2 179,84 84,39 351,63 184,09 12,66

Y3 102 0.51 88,12 462,43 0,65

Drusen

BIOIMAGING2014-InternationalConferenceonBioimaging

Visual inspection of the images showed good

compliance between areas of highest intensity pixels

in first principal component with drusen areas

marked by the experts in original images.

4 DISCUSSION

Similarity of values in the first eigenvector used to

construct first principal component to the color

coefficients proposed in (Patasius 2009-2) shows

that method constructs particular color scheme

similar to the universal one. However we expect that

our scheme will compensate influence of registering

conditions and other technical factors eventually

having critical impact on final result. This advantage

will be proven in future experiments.

The analyzed raw data are homologous in all

initial dimensions because of the same origin (the

same type of registering equipment only in different

colors). Therefore is no need to perform any

normalization and we can’t expect any better results

from higher-level multivariate methods (e.g. Kernel

PCA).

Usage of certain principal component instead of

original image can increase the performance of

automatic drusen area evaluation algorithms. Visual

evaluation of this principal component can reveal

more image details for the expert.

The usage of the method is not limited to the

drusen. It could be used to increase contrast of other

unicolor objects in the images as well.

Limitations of the method: We have only three

original variables determining limited space for

calculated principal components. So one should be

sure that part of the variance corresponding to

drusen should be at least amongst top three,

otherwise PCA will ignore it. It means that some

critical minimal area of the image should be covered

by the drusen, exact percentage of it will be

determined in further investigations.

5 CONCLUSIONS

Principal component analysis based eye fundus

image preprocessing is significantly increasing

contrast-to-noise ratio of drusen area for further

automatic detection or visual examination.

REFERENCES

Bird A. C., Bressler N. M., Bressler S. B., et al. An

international classification and grading system for age-

related maculopathy and age-related macular

degeneration. The International ARM Epidemiological

Study Group. Surv Ophthalmol 1995;39:367–74.

Curcio C. A., Johnson M., Huang J-D., Rudolf M. Aging,

age-related macular degeneration, and the Response-

to-Retention of apolipoprotein B-containing

lipoproteins. Prog Ret Eye Res 2009;28:393– 422.

Gass, J. D. M. Stereoscopic atlas of macular diseases:

diagnosis and treatment. 4. St. Louis: Mosby; 1997.

Jolliffe I. T. Principal Component Analysis. Springer-

Verlag; 2nd edition (2002) ISBN: 0387954422

Patašius M., Marozas V., Jegelevičius D., Lukoševičius A.

Optimal Combinations of Color Space Components

for Detection of Blood Vessels in Eye Fundus Images.

Electronics And Electrical Engineering 2009. No.

3(91) 53-56.

Patašius M., Marozas V., Jegelevičius D., Lukoševičius

A., Špečkauskas M. “Biomedical Engineering“, Proc.

Int Conf. Kaunas 2009;160:163.

Russell S. R., Mullins R. F., Schneider B. L., Hageman G.

S. Location, substructure, and composition of basal

laminar drusen compared with drusen associated with

aging and age-related macular degeneration. Am J

Ophthalmol 2000;129:205–14. [PubMed: 10682974]

Spaide R. F., Curcio C. A. Drusen characterization with

multimodal imaging. Retina. 2010 Oct;30(9):1441-54.

PrincipalComponentAnalysisbasedColourSchemeOptimisationinEyeFundusImages-ContrastEnhancementfor

DetectionandEvaluationofDruseninAgeRelatedMacularDegenerationPatients'Followup