Feature Selection Applied to Human Tear Film Classiﬁcation

Daniel G. Villaverde

, Beatriz Remeseiro

, Noelia Barreira

, Manuel G. Penedo

and Antonio Mosquera

Department of Computer Science, University of A Coru˜na, Campus de Elvi˜na s/n, 15071 A Coru˜na, Spain

Department of Electronics and Computer Science, University of Santiago de Compostela, Campus Universitario Sur,

15782 Santiago de Compostela, Spain

Keywords:

Tear Film, Dry Eye Syndrome, Color Texture Analysis, Feature Selection, Filter Methods, Machine Learning.

Abstract:

Dry eye is a common disease which affects a large portion of the population and harms their routine activities.

Its diagnosis and monitoring require a battery of tests, each designed for different aspects. One of these

clinical tests measures the quality of the tear ﬁlm and is based on its appearance, which can be observed

using the Doane interferometer. The manual process done by experts consists of classifying the interferometry

images into one of the ﬁve categories considered. The variability existing in these images makes necessary

the use of an automatic system for supporting dry eye diagnosis. In this research, a methodology to perform

this classiﬁcation automatically is presented. This methodology includes a color and texture analysis of the

images, and also the use of feature selection methods to reduce image processing time. The effectiveness of

the proposed methodology was demonstrated since it provides unbiased results with classiﬁcation errors lower

than 9%. Additionally, it saves time for experts and can work in real-time for clinical purposes.

1 INTRODUCTION

Dry eye syndrome, resulting from an inadequate tear

ﬁlm, is a prevalent disease which affects a wide range

of population. It affects over 14% of 65+ age group

according to one US study (Moss, 2000), and over

30% of the same group in a population of Chinese

subjects (Jie et al., 2008). The percentage of Euro-

pean people affected by dry eye is quite similar. In

Germany, for example, one in four patients consulting

an ophthalmologist complains of the symptoms of dry

eye (Brewitt and Sistani, 2001). Many sufferers will

require treatment and the potential cost is signiﬁcant.

Therefore, monitoring the effect of the different treat-

ments is of great importance in order to ensure the

maximum beneﬁt to each patient (Bron, 2001).

The tear ﬁlm is a thin ﬁlm formed by the tear ﬂuid

over the exposed ocular surface. It was classically de-

ﬁned by Wolff (E.Wolff, 1954) as a multi-layer struc-

ture which consists of an outer lipid layer, an inter-

mediate aqueous layer, and a deep mucous layer. One

aspect of tear ﬁlm assessment is the evaluation of

the lipid superﬁcial layer (Guillon and Guillon, 1997;

Thai et al., 2004), since it plays an important role in

the retention of the tear ﬁlm by slowing evaporation.

A deﬁciency of this layer may cause evaporative dry

eye syndrome. Although this layer is transparent, in-

terference fringes are created when light rays, reﬂect-

ing off the previous surface, interfere with rays which

have been reﬂected by the posterior surface (Freeman

and Hull, 2003). In this manner, both light and dark

bands of interference can be observed. The number

and spacing of these bands depend on the thickness

of the lipid layer and the rate at which this changes

(Freeman and Hull, 2003).

The Doane tear ﬁlm video interferometer (Doane,

1989) consists in a light source and an observation

system which captured the appearance of the tear ﬁlm

using a video-based system. Using this interferome-

ter, Thai et al. (Thai et al., 2004) measured the evap-

oration rate, thinning characteristics and lipid layer

changes in the tear ﬁlm. For this purpose, they pro-

posed a grading system in order to classify tear ﬁlm

images into different categories.

Based on (Thai et al., 2004), a new grading scale

composed of ﬁve categories was proposed in (Reme-

seiro et al., nd), since the use of a digital camera pro-

duced changes in the detail seen in the digital images.

The variability in appearance of these categories re-

sulted in a major intra- as well as inter- observer varia-

tions, and so a computer-based analysis was also pre-

sented in (Remeseiro et al., nd). This previous re-

395

G. Villaverde D., Remeseiro B., Barreira N., G. Penedo M. and Mosquera A..

Feature Selection Applied to Human Tear Film Classiﬁcation.

DOI: 10.5220/0004809403950402

In Proceedings of the 6th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2014), pages 395-402

ISBN: 978-989-758-015-4

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

search demonstrated that the interference phenomena

can be characterized as color texture patterns, and so

the manual process can be automated with the beneﬁts

of being faster and unaffectedby subjective factors. A

wide set of texture analysis techniques, color spaces

and machine learning algorithms were analyzed. As a

summary, the co-occurrence features method for tex-

ture extraction and the Lab color space produced the

best results with maximum accuracy over 90%. Re-

garding the machine learning algorithms, an statisti-

cal comparison of classiﬁers was performed and the

SVM was established as the most competitive method

for the problem at hand. To the best knowledge of the

authors, there are no more attempts in the literature to

automatically analyze tear ﬁlm images acquired with

the Doane interferometer.

The problem with the approach proposed in

(Remeseiro et al., nd) is that the time required to ex-

tract some of the texture features is too long for clin-

ical purposes. Reducing processing time is a criti-

cal issue in this application, so feature selection tech-

niques will be applied in this research in order to de-

crease the number of features and, as a consequence,

the computational time. The proposed methodology

will provide reliable results in a short period of time

and so the system would be highly recommended for

clinical use in dry eye diagnosis.

This paper is organized as follows: Section 2 de-

scribes the steps of the methodology, Section 3 ex-

plains the experimental study performed, Section 4

shows the results and discussion, and Section 5 in-

cludes the conclusions and future lines of research.

2 RESEARCH METHODOLOGY

The methodology proposed in this research is illus-

trated in Figure 1, and aims at improving automatic

human tear ﬁlm classiﬁcation. First, image process-

ing is performed in order to obtain the quantitative

vectors with the texture and the color information of

the images. Second, feature selection methods are

applied in order to select the subset of relevant fea-

tures. Then, the classiﬁcation step is performed and,

ﬁnally, two performance measures are computed to

evaluate the effectiveness of the methodology. Next,

every stage will be explained in detail.

 

Image

processing

Feature

selection

Classification

Performance

evaluation

Figure 1: Steps of the research methodology.

2.1 Image Processing

The ﬁrst step consists in processing an input image

acquired with the Doane interferometer and, as a re-

sult, a quantitative vector of features is obtained. It is

composed of three sub-steps: (1) the region of interest

of an input image is extracted; (2) this region in RGB

is converted to a speciﬁed color space; (3) each single

channel of the transformed image is analyzed in terms

of texture. As a result, a feature vector is generated.

In what follows, every step will be explained in depth.

2.1.1 Extraction of the Region of Interest

The input images, as depicted in Figure 2(a), include

an external area that does not contain useful infor-

mation. Also, the most relevant information appears

in the central part of the yellowish or greenish area,

formed by the anterior surface of the tear ﬁlm. This

forces a pre-processing step aimed at extracting the

region of interest (ROI) (Remeseiro et al., nd).

(a) (b)

Figure 2: (a) Input image in RGB. (b) Region of interest.

The relevant region of the image is characterized

by green or yellow tonalities, and so only the green

channel of the input image in RGB is considered.

The backgroundof the image is determined by ﬁnding

those pixels whose gray level is less than a threshold:

th = µ − p× σ (1)

where µ is the mean value of the gray levels of the

image, σ is its standard deviation and p is a weight

factor empirically determined.

Once the background is identiﬁed, the central part

of the rest of the image can be located in order to

extract the ﬁnal ROI (see Figure 2(b)). Since some

images include other irrelevant regions, such as eye-

lashes, the morphological operator of erosion (Gonza-

lez and Woods, 2008) is applied in order to eliminate

them from further analysis. Next, a rectangle within

the region identiﬁed above is selected and reduced by

a pre-speciﬁed percentage. This region is likely to be

free of irrelevant features and so, in most of cases,

corresponds to the ﬁnal ROI. However, a ﬁnal step is

needed in some cases: an iterative process to reduce

the size of the ROI until no background areas remain.

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

396

2.1.2 Color Analysis

Some categories show different color characteristics

and, for this reason, images were analyzed not only in

grayscale but also using the Lab color space.

• A grayscale image represents the color as gray

levels from black to white. In order to analyze the

texture over grayscale images, the three channels

of the ROI in RGB are converted into a single gray

channel and its texture is subsequently extracted.

• The CIE 1976 L*a*b color space (McLaren,

1976) is a chromatic color space appropriate for

texture analysis. Using the Lab color space entails

converting the three channels of the ROI in RGB

to the three components of Lab. Then, each com-

ponent is individually analyzed in terms of tex-

ture and the ﬁnal descriptor is the concatenation

of these three individual descriptors.

2.1.3 Texture Analysis

Texture is used to characterize the interference pat-

terns of the grading scale considered. As several

methods for texture analysis could be applied, ﬁve

popular techniques are tested in this research:

• Butterworth ﬁlters (Gonzalez and Woods, 2008)

are frequency domain ﬁlters that have an approx-

imately ﬂat response in the bandpass frequency,

which gradually decays in the stopband. The or-

der of the ﬁlter deﬁnes the slope of the decay. A

bank of Butterworth ﬁlters composed of 9 second

order ﬁlters was used, with bandpass frequencies

covering the whole frequency spectrum. The ﬁl-

ter bank maps each input image to 9 output im-

ages, one per frequency band. Each output image

was normalized separately and then a uniform his-

togram with non-equidistant bins was computed.

Since 16-bin histograms were used, the feature

vectors contain 16 components per ﬁlter.

• Gabor ﬁlters (Gabor, 1946) are complex expo-

nential signals modulated by Gaussian functions.

The parameters of Gabor ﬁlters deﬁne their shape,

and represent their location in the spatial and fre-

quency domains. A bank of 16 Gabor ﬁlters cen-

tered at 4 frequencies and 4 orientations was cre-

ated. The ﬁlter bank maps each input image to 16

output images, one per frequency-orientation pair.

Using the same idea as in Butterworth ﬁlters, the

descriptor of each output image is its uniform his-

togram with non-equidistant bins. Since 13-bin

histograms were used in grayscale and 17-bin his-

tograms were used in Lab, the feature vectors con-

tain 13 and 17 components per ﬁlter, respectively.

• The discrete wavelet transform (Mallat, 1989)

generates a set of wavelets by scaling and translat-

ing a mother wavelet. The wavelet decomposition

of an image consists of applying these wavelets

horizontally and vertically, generating 4 subim-

ages (LL, LH, HL, HH) which are then subsam-

pled by a factor of 2. Then, the process is repeated

n− 1 times over the LL subimage, where n is the

number of scales. A generalized Daubechies al-

gorithm (Daubechies, 1992) was applied as the

mother wavelet. The descriptor of an input image

is constructed by calculating the mean and the ab-

solute average deviation of the input and LL im-

ages, and the energy of the LH, HL and HH im-

ages. Since 8 scales were used, the feature vectors

contain 42 components.

• Markov random ﬁelds generate a texture model by

expressing the gray values of each pixel in an im-

age as a function of the gray values in a neighbor-

hood of the pixel. In order to generate the descrip-

tor, the directional variances proposed in (C¸ esmeli

and Wang, 2001) were used. In this work, the

neighborhood of a pixel is deﬁned as the set of

pixels within a Chebyshev distance d. Distances

from 1 to 10 were considered and, for a distance

d, the descriptor comprises 4d features.

• Co-occurrence features analysis (Haralick et al.,

1973) is based on the computation of the condi-

tional joint probabilities of all pairwise combina-

tions of gray levels. For a given distance and an

orientation, this method generates a set of gray

level co-occurrence matrices and extracts several

statistics from their elements. In general, the num-

ber of orientations and matrices for a distance d

is 4d. A set of 14 statistics proposed in (Har-

alick et al., 1973) was computed from each co-

occurrence matrix. The descriptor of an image

consists of 2 properties,the mean and range across

matrices of these statistics, thus obtaining a fea-

ture vector with 28 components per distance. Us-

ing the Chebyshev distance as in Markov random

ﬁelds, distances from 1 to 17 were considered.

2.2 Feature Selection

Feature selection is a dimensionality reduction tech-

nique which consists of removing the irrelevant and

redundant features in order to obtain a reduction in

processing time without a degradation in performance

(Guyon et al., 2006). Among the different feature se-

lection techniques that can be found in the literature,

ﬁlters are used in this research for several reasons: (1)

they are the least time-consuming, (2) they rely on

the general characteristics of the training data, and (3)

FeatureSelectionAppliedtoHumanTearFilmClassification

397

they are independent of the induction algorithm. In

summary, ﬁlters are computationally simple and fast.

Three ﬁlters were selected in this research based on

previous works (Bol´on-Canedo et al., 2013):

• Correlation-based feature selection (Hall, 1999)

is a simple algorithm that ranks feature subsets ac-

cording to a correlation based heuristic. The bias

of this function is toward subsets that contain fea-

tures that are highly correlated with the class and

uncorrelated with each other. Irrelevant features

should be ignored because they will have low cor-

relation with the class; whilst redundant features

should be discarded since they will be highly cor-

related with at least one of the remaining features.

• Consistency-based ﬁlter (Dash and Liu, 2003)

evaluates the worth of a subset of features by

the level of consistency in the class values when

the instances are projected onto the subset of at-

tributes. The algorithm generates a random subset

S from the number of features in every round. If

the number of features of S is lower than the best

current set (S

best

), the data with the features pre-

scribed in S is checked against the inconsistency

criterion. If its inconsistency rate is below a pre-

speciﬁed one, S becomes the new S

best

. The in-

consistency criterion speciﬁes to what extent the

dimensionally reduced data can be accepted.

• INTERACT (Zhao and Liu, 2007) is a subset ﬁl-

ter based on symmetrical uncertainty (SU) and the

consistency contribution (CC), which is an indi-

cator about how signiﬁcantly the elimination of a

feature will affect consistency. The algorithm is

made up of two parts: (1) the features are ranked

in descending order based on their SU values, and

(2) the features are evaluated one by one starting

from the end of the ranked feature list. If the CC

of a feature is lower than an established threshold,

the feature is removed, otherwise it is selected.

2.3 Classiﬁcation

Once all the features are extracted from the ROI of a

single image, the obtained descriptor has to be clas-

siﬁed into one of the considered categories. For this

task, a support vector machine (SVM) (Burges, 1998)

is trained according to the results presented in (Reme-

seiro et al., nd). In the aforesaid paper, an statisti-

cal analysis of machine learning algorithms was per-

formed, and the SVM was selected as the most com-

petitive method for the problem at hand, compared

with methods such as Naive Bayes or decision trees.

2.4 Performance Evaluation

After the SVM is trained, the performance of the sys-

tem is evaluated in terms of two different measures of

relevance to the problem in question:

• The classiﬁcation error is computed as the per-

centage of incorrectly classiﬁed instances.

• The feature extraction time is computed as the

time that the texture analysis methods take to ex-

tract the selected features from a single image.

Note that this does include neither the time of

training a classiﬁer nor the time of performing

feature selection, since they are off-line processes.

3 EXPERIMENTAL STUDY

The aim of this research is to present a methodology

based on color texture analysis and feature selection

to classify tear ﬁlm images acquired with the Doane

interferometer. This methodology is tested in order to

improve previous results (Remeseiro et al., nd). The

materials and methods used in this research are pre-

sented in this section.

3.1 Data Source

In order to test the proposed methodology, a bank of

images acquired from dry eye patients with average

age 55 ± 16 was used. This dataset is publicly avail-

able in (VOPTICAL

GCU, nd). All images in this

bank have been annotated by two optometrists from

the Department of Life Sciences, Glasgow Caledo-

nian University (Glasgow, UK).

The acquisition of the images was carried out with

the Doane interferometer (Doane, 1989) and a digi-

tal PC-attached CMEX-1301 camera (CMEX-1301x,

nd). The ImageFocus Capture and Analysis soft-

ware (ImageFocus, nd) was used for image capture,

and images were stored at a spatial resolution of

1280× 1024 pixels in the RGB color space. Multiple

images were taken for up to one minute and, due to the

various artifacts associated with image capture, many

of them were unsuitable for analysis. Therefore, op-

tometrists selected only those images taken shortly af-

ter blinking, and when the eye was fully open.

The bank is composed of 106 images and in-

cludes samples from the ﬁve categories considered:

11 strong fringes, 25 coalescing strong fringes, 30

ﬁne fringes, 26 coalescing ﬁne fringes and 14 debris.

These grades of interference patterns were deﬁned by

experimented optometrists in (Remeseiro et al., nd),

and are deﬁned as follows (see Figure 3):

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

398

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

Figure 3: (a) Strong fringes. (b) Coalescing strong fringes. (c) Fine fringes. (d) Coalescing ﬁne fringes. (e) Debris.

• Strong fringes. Obvious color fringes with an ap-

pearance of spreading across the cornea.

• Coalescing strong fringes. Obvious color fringes,

but coalescing into islands of color.

• Fine fringes. Gray fringes with an appearance of

spreading across the cornea.

• Coalescing ﬁne fringes. Gray fringes, but coalesc-

ing into islands of gray.

• Debris. Obvious disturbances in the tear ﬁlm,

likely to be of varying origin.

3.2 Experimental Procedure

The experimental procedure is detailed as follows:

• Apply the ﬁve texture analysis methods (see Sec-

tion 2.1.3) to the dataset of images using the two

color spaces (see Section 2.1.2).

• Apply the three feature selection ﬁlters (see Sec-

tion 2.2) to the dataset to obtain the subset of fea-

tures that properly describe the problem at hand.

• Train a SVM (see Section 2.3) with radial ba-

sis kernel and automatic parameter estimation. A

leave-one-out cross-validation was used, so the

average error across all trials was computed.

• Evaluate the effectiveness of the proposed

methodology in terms of two performance mea-

sures (see Section 2.4).

Experimentation was performed on an

Intel

Core

i5 CPU 760 @ 2.80GHz with

RAM 4 GB.

4 RESULTS AND DISCUSSION

The results obtained with each color space (grayscale

and Lab), each texture analysis method (Butterworth

ﬁlters - BF, Gabor ﬁlters - GF, the discrete wavelet

transform - DWT, Markov random ﬁelds - MRF, and

co-occurrence features - CF), and each feature selec-

tion ﬁlter (correlation-based feature selection - CFS,

consistency-based ﬁlter - Cons, and INTERACT -

INT) will be analyzed in terms of the two perfor-

mance measures described above (classiﬁcation error

and feature extraction time). Bear in mind that the

column None in the tables of this section shows the re-

sults when no feature selection was performed. Also,

as a leave-one-out cross-validation was performed,

the error, the number of features, and the time for ex-

tracting them were computed as an average value.

Table 1 shows the number of features selected by

each of three feature selection ﬁlters. In average, CFS

and INTERACT retain the 4.9% and 2.08% of the

features, respectively; whilst consistency-based ﬁl-

ter performed the most aggressive selection retaining

only the 1.4% of the features.

Table 1: Number of features. From top to bottom, each cell

contains the results corresponding to grayscale and Lab.

None CFS Cons INT

144 12.52 8.91 11.43

432 20.65 5.89 12.75

208 13.92 7.06 9.68

816 23.69 6.28 12.72

DWT

42 10.95 5.9 6.18

126 17.67 6.05 8.80

MRF

220 9 7.23 8.87

660 21.34 6.42 11.56

476 37.64 4.49 16.83

1428 55.9 4.3 28.75

Table 2 shows the test errors for all color spaces,

texture analysis methods and feature selection ﬁlters

after applying the SVM classiﬁer, where the best re-

sult for each combination appears highlighted. In

general, all texture analysis techniques perform quite

well providing results below 16% error. Regarding

the texture and color analysis without performing fea-

ture selection, it can be seen that the use of color in-

formation slightly improves the results achieved when

compared to grayscale analysis since some categories

contain not only morphological features, but also

color features. On the other hand, all texture extrac-

tion methods perform quite well, but co-occurrence

features generate the best results, closely followed

by Gabor ﬁlters. Despite the fact that Markov ran-

dom ﬁelds use neighborhood information, as does

FeatureSelectionAppliedtoHumanTearFilmClassification

399

co-occurrence features analysis, the Markov method

does not perform as well because less information is

included in the ﬁnal descriptor. In essence, the combi-

nation of co-occurrence features and the INTERACT

ﬁlter outperform the other methods, with the best re-

sult of 9.4% error using grayscale images. Regard-

ing feature selection, it outperforms primal results in

ﬁve out twelve pairwise combinations of color spaces

and texture analysis techniques using CFS and IN-

TERACT ﬁlters. However, the consistency-based ﬁl-

ter produced a degradation in performance in all the

combinations due to its aggressive selection.

Table 2: Mean test classiﬁcation error (%). From top

to bottom, each cell contains the results corresponding to

grayscale and Lab.

None CFS Cons INT

15.10 18.87 18.87 20.76

16.88 21.70 18.87 15.10

16.04 25.47 28.31 20.76

11.32 16.98 33.02 17.93

DWT

24.53 20.76 30.19 20.76

21.70 13.21 18.87 20.76

MRF

21.70 22.65 26.42 23.59

15.10 24.53 23.59 25.47

11.32 11.32 13.21 9.44

11.32 13.21 16.04 15.10

The automatic grading system for human tear

ﬁlms should provide results to the clinicians in a very

brief period of time, since waiting too long in front

of a computer could be a reason for rejection of its

use. In this sense, applying feature selection meth-

ods to reduce the number of input attributes and, thus,

the time needed for extracting texture features can be

a key step in order to improve the automatic system.

In this sense, Table 3 shows the times needed for an-

alyzing color and texture information of one single

image, where the best result for each combination ap-

pears highlighted. Notice that co-occurrence features

has been known to be slow and, for this reason, an op-

timization of the method based on (Clausi and Jerni-

gan, 1998) was used in this research.

According to Tables 2 and 3, the effectiveness of

using feature selection is demonstrated since, in most

cases, the time is signiﬁcantly reduced without wors-

ening the performance. In eleven out twelve pairwise

combinations, the lowest times are obtained using

the most aggressive algorithm, which is consistency-

based ﬁlter, and in only one case INTERACT outper-

forms it. The maximum processing time that would

be accepted in this clinical system is around 10 or 20

seconds, and only two methods for texture analysis

can be rejected for this reason (Butterworth ﬁlters and

Markov random ﬁelds). Furthermore, both discrete

Table 3: Feature extraction time (s). From top to bottom,

each cell contains the results corresponding to grayscale and

Lab.

None CFS Cons INT

101.76 68.25 62.82 63.89

305.17 132.96 47.96 109.60

20.72 8.32 6.80 7.49

62.70 19.89 8.09 13.38

DWT

0.96 0.68 0.46 0.44

3.00 1.45 0.59 0.89

MRF

74.05 50.82 33.20 56.98

221.72 104.96 29.29 67.02

58.69 6.22 0.42 3.31

169.22 20.58 0.34 8.47

wavelet transform and co-occurrence features analy-

sis take less than 1 second in providing results so the

system could work in real-time using these two tex-

ture analysis techniques.

Among all the combinations of methods, it is a

very difﬁcult task to select the best combination since

two performance measures are considered. Although

co-occurrence features analysis outperforms the other

methods in terms of classiﬁcation error and provides

good results regarding feature extraction time, the

Pareto front (Teich, 2001) for each alternative was

computed in order to analyze the balance between er-

ror and time. In the context of multi-objective op-

timization, the Pareto front is deﬁned as the set of

points which are equally satisfying the constraints of

the corresponding problem. Thus, selecting a solution

in the Pareto front would imply to select a better solu-

tion than any outside the Pareto front. In this research,

solutions are constrained to minimize both classiﬁca-

tion error and feature extraction time. Figure 4 shows

the points in the Pareto front marked with a red circle.

These three points correspond to: (1) co-occurrence

features using grayscale images and INTERACT, (2)

co-occurrence features using grayscale images and

consistency-based ﬁlter, and (3) co-occurrence fea-

tures using images in the Lab color space after ap-

plying the consistency-based ﬁlter. The selection of

one of these three points will depend on if it is prefer-

able to minimize either the error or the time for the

problem at hand.

The three points in the Pareto front were analyzed

in order to try to shed light on this issue. For rea-

sons of simplicity, these three points will be referred

with the numbers in Figure 4. Using the combina-

tion (2), the number of features considered is ap-

proximately the same than using the combination (3)

(4.49 and 4.3 features in average). However, the time

needed for extracting these features is lower if the op-

tion (3) is considered (0.42 and 0.34 seconds in aver-

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

400

Figure 4: Pareto front for all the combinations of methods

for texture extraction, color analysis and feature selection.

age). Comparing the options (1) and (2), it can be

seen that the number of features in (1) is almost 4

times the number of features in (2), whilst the time

is 8 times greater. This fact lead us to a depth anal-

ysis of the features extracted with the co-occurrence

features analysis, since the time needed for obtaining

its features is not homogeneous. Features are vector-

ized in groups of 28 related to distances and channels

in the color space. Each group of 28 features corre-

sponds with the mean and range of 14 statistics across

the co-occurrence matrices. Some experiments were

performed on the time the method takes to compute

each of the 14 statistics. Results disclosed that com-

puting only the 14

statistic, which corresponds with

the maximal correlation coefﬁcient (Haralick et al.,

1973), takes around 60% of the total time.

After this analysis focus on the individual features

of the method proposed by Haralick, the aim here is to

explore the impact of removing all the 14

statistics.

In this manner, Tables 4 and 5 show the classiﬁcation

error and the feature extraction time, respectively, for

this experiment. From top to bottom, each cell con-

tains the results previously obtained and the results if

the 14

statistics are excluded from each set of fea-

tures. As can be seen in Table 4, the error is reducedin

three out eight combinations, and maintained in two

out eight combinations. Regarding Table 5, the time

is reduced in seven out eight combinations and main-

tained in the other one. Thus, the exclusion of the

statistic allows to improve the performance of the

methodology. However, it is also important to remark

the effectiveness of the INTERACT ﬁlter for selecting

the most appropriate features since simply removing

the 14

statistics from the original set of features not

only takes more time to perform feature extraction,

but also produces worse results in terms of error.

Table 4: Mean test classiﬁcation error (%). From top to

bottom, each cell contains the results corresponding to the

experiment with and without the 14

statistic.

None CFS Cons INT

CF+gray

11.32 11.32 13.21 9.44

15.10 14.15 13.21 8.49

CF+lab

11.32 13.21 16.04 15.10

11.32 10.38 16.98 14.15

Table 5: Feature extraction time (s). From top to bottom,

each cell contains the results corresponding to the experi-

ment with and without the 14

statistic.

None CFS Cons INT

CF+gray

58.69 6.22 0.42 3.31

24.40 3.93 0.42 1.92

CF+lab

169.22 20.58 0.34 8.47

75.59 5.26 0.31 2.71

5 CONCLUSIONS AND FUTURE

RESEARCH

An automatic grading system to measure the quality

of human tear ﬁlms was developed in previous re-

search, but it requires a too long time for extracting

some texture features. This time prevents the clini-

cal use of the system, so the aim of this work is im-

proving previous results focus on reducing the pro-

cessing time. For this task, three of the most popu-

lar feature selection ﬁlters (correlation-based feature

selection, consistency-based ﬁlter and INTERACT)

were tested. Results obtained with this methodology

surpass the previous approach in terms of processing

time and, furthermore, improves slightly the accuracy

of the system.

In clinical terms, the importance of the proposed

methodology lies in providing objective results in

real-time, which saves time for experts who do the

process by hand. Speciﬁcally, the system is able to

automatically classify the images obtained using the

Doane interferometer with an error lower 9% and a

processing time under one second.

Future work will involve performing local analy-

sis in order to segment one single tear ﬁlm image in

different categories. The motivation is that although

some tear ﬁlm images, from an individual subject,

conformed to a single pattern, it was more common

for them to be made up of a combination of differ-

ent patterns. In addition, investigation of dynamic

changes seen in the tear ﬁlm during the inter-blink

time interval would help in identifying subjects with

poor tear ﬁlm stability.

FeatureSelectionAppliedtoHumanTearFilmClassification

401

ACKNOWLEDGEMENTS

This research has been partially funded by the Sec-

retar´ıa de Estado de Investigaci´on of the Spanish

Government and FEDER funds of the European

Union through the research projects PI10/00578 and

TIN2011-25476. Beatriz Remeseiro acknowledges

the support of Xunta de Galicia under Plan I2C Grant

Program.

We would also like to thank the School of Health

and Life Sciences, Glasgow Caledonian University

for providing us with the annotated image datasets.

REFERENCES

Bol´on-Canedo, V., S´anchez-Maro˜no, N., and Alonso-

Betanzos, A. (2013). A review of feature selection

methods on synthetic data. Knowledge and Informa-

tion Systems, 34(3):483–519.

Brewitt, H. and Sistani, F. (2001). Dry Eye Disease: The

Scale of the Problem. Survey of Ophthalmology,

45(2):199–202.

Bron, A. J. (2001). Diagnosis of Dry Eye. Survey of Oph-

thalmology, 45(2).

Burges, C. J. (1998). A Tutorial on Support Vector Ma-

chines for Pattern Recognition. Data Mining and

Knowledge Discovery, 2:121–167.

C¸esmeli, E. and Wang, D. (2001). Texture Segmentation

Using Gaussian-Markov Random Fields and Neural

Oscillator Networks. IEEE Transactions on Neural

Networks, 12.

Clausi, D. and Jernigan, M. (1998). A Fast Method to Deter-

mine Co-occurrence Texture Features. IEEE Transac-

tions on Geoscience and Remote Sensing, 36(1):298–

300.

CMEX-1301x (n.d.). CMEX-1301x camera. Euromex Mi-

croscopen BV, Arnhem, The Netherlands.

Dash, M. and Liu, H. (2003). Artiﬁcial intelligence, 151(1-

2):155–176.

Daubechies, I. (1992). Ten Lectures on Wavelets. SIAM,

CBMS series.

Doane, M. G. (1989). An instrument for in vivo tear

ﬁlm interferometry. Optometry and Vision Science,

66(6):383–388.

E.Wolff (1954). Anatomy of the eye and orbit (4th edition).

H. K. Lewis and Co., London.

Freeman, M. H. and Hull, C. C. (2003). Interference and

optical ﬁlms. Butterworth Heinemann.

Gabor, D. (1946). Theory of Communication. Journal of

Institute for Electrical Engineering, 93:429–457.

Gonzalez, R. and Woods, R. (2008). Digital image process-

ing. Pearson/Prentice Hall.

Guillon, J. P. and Guillon, M. (1997). Tearscope plus clini-

cal hand book and tearscope plus instructions. Keeler

Ltd. Windsor, Berkshire, Keeler Inc, Broomall, PA.

Guyon, I., Gunn, S., Nikravesh, M., and Zadeh, L. (2006).

Feature Extraction: Foundations and Applications.

Springer Verlag.

Hall, M. (1999). Ph.D dissertation, The University of

Waikato.

Haralick, R. M., Shanmugam, K., and Dinstein, I. (1973).

Texture Features for Image Classiﬁcation. IEEE

Transactions on Systems, Man, and Cybernetics In

Systems, Man and Cybernetics, 3:610–621.

ImageFocus (n.d.). ImageFocus Capture and Analysis

software. Euromex Microscopen BV, Arnhem, The

Netherlands.

Jie, Y., Xu, L., Wu, Y. Y., and Jonas, J. B. (2008). Preva-

lence of dry eye among adult Chinese in the Beijing

Eye Study. Eye, 23(3):688–693.

Mallat, S. G. (1989). A theory for multiresolution signal de-

composition: the wavelet representation. IEEE Trans-

actions on Pattern Analysis and Machine Intelligence,

11:674–693.

McLaren, K. (1976). The development of the CIE 1976

(L*a*b) uniform colour-space and colour-difference

formula. Journal of the Society of Dyers and

Colourists, 92(9):338–341.

Moss, S. E. (2000). Prevalence of and Risk Factors for

Dry Eye Syndrome. Archives of Ophthalmology,

118(9):1264–1268.

Remeseiro, B., Oliver, K., Tomlinson, A., Martin, E., Bar-

reira, N., and Mosquera, A. (n.d.). Automatic grading

system for human tear ﬁlms. Under review.

Teich, J. (2001). Pareto-front exploration with uncertain ob-

jectives. In Evolutionary multi-criterion optimization,

volume 1993, pages 314–328. Springer.

Thai, L. C., Tomlinson, A., and Doane, M. G. (2004). Effect

of Contact Lens Materials on Tear Physiology. Op-

tometry and Vision Science, 81(3):194–204.

VOPTICAL

GCU (n.d.). VOPTICAL GCU, VARPA

optical dataset annotated by optometrists from

the Department of Life Sciences, Glasgow Cale-

donian University (UK). [Online] Available:

http://www.varpa.es/voptical

gcu.html, last access:

december 2013.

Zhao, Z. and Liu, H. (2007). Searching for interacting fea-

tures. pages 1156–1161.

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

402