IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 17, NO.
1, JANUARY 2013
Iris-Based Medical Analysis by Geometric
Deformation Features
Lin Ma, David Zhang, Fellow, IEEE, Naimin Li, Yan Cai, Wangmeng Zuo, Member, IEEE,
and Kuanquan Wang, Senior Member, IEEE
Abstract—Iris analysis studies the relationship between human
health and changes in the anatomy of the iris. Apart from the fact
that iris recognition focuses on modeling the overall structure of
the iris, iris diagnosis emphasizes the detecting and analyzing of lo-
cal variations in the characteristics of irises. This paper focuses on
studying the geometrical structure changes in irises that are caused
by gastrointestinal diseases, and on measuring the observable
deformations in the geometrical structures of irises that are re-
lated to roundness, diameter, and other geometric forms of the
pupil and the collarette. Pupil- and collarette-based features are
defined and extracted. A series of experiments are implemented
on our experimental pathological iris database, including manual
clustering of both normal and pathological iris images, manual
classification by nonspecialists, manual classification by individ-
uals with a medical background, classification ability verification
for the proposed features, and disease recognition by applying the
proposed features. The results prove the effectiveness and clini-
cal diagnostic significance of the proposed features and a reliable
recognition performance for automatic disease diagnosis. Our re-
search results offer a novel systematic perspective for iridology
studies and promote the progress of both theoretical and practical
work in iris diagnosis.
Index Terms—Collarette-based feature, disease recognition, iris
geometric deformation feature, pupil-based feature.
I. INTRODUCTION
IOMETRIC methods identify people or analyze the status
B of human health based on physical or behavioral charac-
teristics [1], [2]. Most biometric methods measure human body
surface characteristics, such as finger prints, palm prints, face,
irises, tongue, breath, etc. [2]. Computer-aided biometrics has
Manuscript received February 5, 2012; revised July 9, 2012; accepted
September 24, 2012. Date of current version February 4, 2013. This work was
supported in part by the Specialty Leader Support Foundation of Harbin under
Grant 2003AFLXJ006, in part by the Foundation of Heilongjiang Province for
Scholars Returned from Abroad under Grant LC04C17, in part by the Fun-
damental Research Funds for the Central Universities HIT.NSRIF.2012047,
and in part by the National Natural Science Foundation of China under Grant
61271345.
L. Ma, N. Li, W. Zuo, and K. Wang are with the School of Computer Science
and Technology, Harbin Institute of Technology, Harbin 150001, China (e-mail:
malin_li@hit.edu.cn).
D. Zhang is with the Shenzhen Graduate School, Harbin Institute of Tech-
nology, Shenzhen 518055, China, and also with the Department of Computing,
Biometrics Research Centre, Hong Kong Polytechnic University, Kowloon,
Hong Kong (e-mail: csdzhang@comp.polyu.edu.hk).
Y. Cai is with the Fourth Affiliated Hospital, Harbin Medical University,
Harbin 150001, China.
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TITB.2012.2222655
1089-7771/$31.00 © 2012 IEEE
223
made great achievements in recent years. For example, biomet-
ric methods based on the spatial pattern of the iris are considered
to attain a very high accuracy, and therefore, there has been an
explosion of interest in iris biometrics in recent years [3]. At
the same time, people may now diagnose appendicitis based on
detecting tongue color, external characteristics on the tongue
or blood vessels under the tongue [4]; recognize gastritis and
cholecystitis by collecting and automatically analyzing pulse
conditions [5]; and apply breath analyses in medicine and clin-
ical pathology [6]. Increasingly, more attention and demand
have been placed onto such noninvasive, real-time, harmless,
and practical biometric methods to address health issues by
both individuals and hospitals.
The iris is the “colored ring of tissue around the pupil through
which light enters the interior of the eye” [3], [7]. The iris is an
inner body organ, but also has externally visible characteristics.
The fact that an iris contains a large amount of characteristic
information in terms of complex and precise structural and tex-
tural details makes iris recognition and iris diagnosis (iridology)
two important directions in biometrics study.
Biometric methods based on the spatial pattern of the iris are
believed to result in very high accuracies, which have led to an
explosion of interest in iris biometrics in recent years. The main
idea of iris recognition is that the minute details of the iris tex-
ture can be randomly determined during the fetal development
of the eye [8], [9], and the color of the iris can change as the
amount of pigment in the iris increases during childhood [3].
However, this argument undermines the fact that the iris pos-
sesses different kinds of variability, which produces special iris
features. In fact, in order to assure a satisfactory performance
in most applications, iris recognition systems demand a regular
re-enrollment of users. This implies that the iris structure and
surface may be affected by various conditions and changes to
iris codes may be obvious enough to cause mismatchings [10].
Iridology methods based on structural and textural changes in
the local iris areas are regarded as promising in health surveil-
lance and subhealth diagnosis. The main idea of iridology is that
the health status will be reflected on the iris due to two reasons.
First, the anatomical characteristics of the iris bring about
great variability to the iris. There are two muscles which are
called the dilator and the sphincter muscles that control the size
of the iris to adjust for the amount of light that enters the pupil.
Iris tissue does not follow a perfect “rubber sheet” model when
the pupil dilates and contracts [10]. This variation of the iris
surface characteristics becomes a common problem in all iris
recognition methods, because it greatly affects the quality of a
match between iris images [3]. Injuries and indirect influences
224 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 17, NO. 1, JANUARY 2013
caused by disease or health changes are naturally directed to
the dilator, and the sphincter muscles will bring about some
complex nonlinear deformations to the iris.
The iris is composed of five layers of complex connective
and fibrous tissues, and typically has a rich pattern of furrows,
ridges, and pigment spots [9], holes, stripes, and filaments. Any
disease that affects the iris pattern and the proportion and the
uniformity of the collarette will be easily observed on the iris.
The iris is rich in blood vessels. Nerve fibers comprise 70%
of the iris [11]–[13]. The iris nerves are connected to the brain
central and the autonomic nervous systems. The anatomical
characteristics of the iris will be reflected by the body’s health
status and functional changes in the organs through the nervous
system, the blood circulation system, and hormone changes [12],
[13].
In general, the variability of the iris not only influences the
performance of iris recognition, but also constructs the anatom-
ical basis for iris diagnosis.
Second, many inherent and most of the acquired medical
conditions may affect iris biometrics. Caused by the deletion on
a chromosome, aniridia is the most typically inherent abnormal
iris feature [3]. In such a case, the person is born without an iris,
or with a partial iris. Aniridia is estimated to have an incidence
of occurrence between 1 in 50 000 and 1 in 100 000 people [10].
Another inherent iris feature is the color of the iris. Some people
believe that the color of the iris can change as the amount of
pigment in the iris increases during childhood [9]. However,
there are no conclusive results on the age limit of childhood
and the color remains constant during the whole lifespan. On
the contrary, it is almost common knowledge that liver disease
usually brings about changes to the color of the iris.
Apart from the inherent reasons, many common and fre-
quently occurring eye diseases produce obvious phenomenons
to the iris. Aside from inherited aniridia, acquired aniridia is also
often observed in clinical practice, and there are reports on the
discovery of its disease causing genes [14]. Cataracts, a natural
result of aging, are a clouding of the lenses, and greatly affect
the iris zone close to the pupillary boundary. Eye injuries, cer-
tain medications, and diseases such as diabetes and alcoholism
have been known to cause cataracts. About 50% of people aged
65–74 and about 70% of those aged 75 and older have visually
significant cataracts [3], [9]. Cataracts can be removed through
surgery; however, the iris is greatly influenced and people need
to re-enroll in iris biometric systems [3], [15]. Glaucoma, mainly
marked by an increase of pressure inside the eye, causes optic
nerve damage and vision loss. Glaucoma generally occurs with
increased incidence as people age and is more common among
people of African descent, in conjunction with other conditions
such as diabetes [3], [16]. Albinism is a genetic condition that
“results in partial or full absence of pigment (color) from the
skin, hair, and eyes.” Approximately 1 in 17 000 people is af-
fected by albinism [10].
For iridologists, typical and frequently applied iris features
are color, pigment distribution, position of lacunae, and sev-
eral geometric features in shape and dimension of the pupil and
the collarette [1], [2], [9], [13]. One specific claim is that cer-
tain positions on the anterior iris surface are associated with
Fig. 1. Examples of abnormal irises (pathological irises) taken from our clin-
ical iris database. First row from left to right: gastric ulcer iris of a male,
gastroduodenal bleeding iris of a male and a female. Second row from left to
right: gastric cancer iris of a male, and gastric ulcer iris of another male after a
surgical operation.
particular organs and that the presence of a particular feature(s),
that is, open lacuna, closed lacuna, and crypt, at a specific locus
indicates a weakness or a pathological state in the associated
organ [2], [9].
Our study focuses on discovering the correlation between
geometric deformation features of the pupil and collarette and
digestive diseases, such as stomach and intestinal diseases. The
reason for our focus is that digestive diseases are so-called com-
mon and frequently occurring diseases. The geometric defor-
mation features represented in the iris have a relatively large
scope, and are much more objective, easily observed, and widely
admitted. In general cases, the pupil is usually circular, repre-
senting about 1/3 of the iris. The collarette is located about
1.5 mm from the edge of the pupil. The shapes of the pupil
and collarette (roundness and smoothness of boundaries) and
their relative largeness (proportion of the pupil to the iris, and
the proportion of the collarette to the pupil and to the iris) are
important indicators for a number of gastrointestinal diseases.
Fig. 1 gives some iris examples of patients who suffer from
stomach diseases.
This paper intends to provide a promising attempt to use iris
biometric features for clinical pathology applications. This pa-
per is organized as follows. Section II presents a series of manual
experiments that statistically analyzes the pathological signifi-
cance of geometric iris features in order to verify that abnormal
iris geometric features are significant enough to warrant obser-
vation. Section III defines a series of pupil- and collarette-based
features, studies the optimal parameters for feature extraction,
and also considers their classification ability to distinguish be-
tween normal and pathological irises. Section IV describes the
evaluation of iris geometric features and their performances for
common disease diagnoses. Finally, Section V concludes this
paper.
II. GEOMETRICAL STATISTICS OF TYPICAL IRIS SAMPLES
In the first stage of our iris pathological feature study, some
statistical analyses were performed on our clinical iris database.
The aim was to verify if there exist any kinds of obvious iris
features that may be recognized as specific and significant by
humans, and if so, do the features have statistical significance?
MA et al.: IRIS-BASED MEDICAL ANALYSIS BY GEOMETRIC DEFORMATION FEATURES 225

TABLE I
DATA DISTRIBUTION FOR NORMAL AND PATHOLOGICAL IRISES IN THIS PAPER

Three statistical experiments were conducted: a blind manual

clustering of all irises, a manual classification of normal and ab- Fig. 2. Histogram of a number of individuals who named same subsets.
normal irises, and a manual classification of pathological irises.

A. Iris Database Used for Our Research

TABLE II
Our iris image capture device uses a color charge-coupled de- NUMBER OF RESPONDENTS WHO NAMED A SIMILAR SUBSET
vice camera (with 30 times magnification), with an adjustable
focusing system. This equipment can collect 24-bit color im-
ages and avi format videos. During our data collection, in order
to minimize the effects of hand shaking and other imprecise
focusing factors on the quality of image acquisition, we used
avi format video data. Each iris was continuously recorded for
10 seconds in video.
The iris images used are intercepted from the videos with
good quality. All images are saved as 24-bit true color images.
Under the current data acquisition conditions, since it was nearly
impossible to avoid the impact of ambient light and other inter-
ferences, only gray-level images are considered in this study.
Our clinical iris database contains both normal and abnormal
how they subconsciously inducted the geometric features of the
irises. The iris images were taken from student volunteers at the
irises.
Harbin Institute of Technology. A normal iris refers to the iris
The participants were freshman students at our university
image of a healthy individual. To ensure that the irises came from
from several engineering schools. The problem given to them
healthy individuals, we made our selection by asking several
was simply described as follows: “Please observe the given 292
questions about the current health status and family genetic
iris images and divide them into several reasonable subsets as
history of the volunteers.
you see fit. Please note that you may group images into the
An abnormal iris (pathological iris) refers to the iris of a
same subset that you believe to have some sort of common
patient with a clear medical diagnosis. The information was
characteristics, and each subset should contain no less than 3
mainly taken from the No. 211 Military Hospital and the No. 4
images.”
Affiliated Hospital at the Harbin Medical University. Data were
After the task was completed, each respondent was asked
collected with a primary focus on patients that had gastroin-
to give a name (or short description) to each subset. Once the
testinal diseases. Some examples of abnormal irises are given
experiment was completed, each respondent was asked if they
in Fig. 1.
knew anything about iridology. We selected ten students who
Currently, the iris image database has 734 videos of irises
indicated that they did not know anything about iridology and
from 367 healthy individuals and 432 videos from 216 patients
analyzed their clustering results. Sixteen different subsets were
diagnosed with gastrointestinal diseases.
named, and the number of individuals who named the same
In this paper, five types of common and frequently occurring
subsets is listed in a histogram as shown in Fig. 2. Information
gastrointestinal diseases are considered, which are as follows:
on the six largest subsets is given in Table II.
gastrointestinal bleeding, gastric ulcers, enteritis, gastric cancer,
Six primary types of irises were found: 1) ordinary; 2) overly
and intestinal cancer, covering a total number of 142 people. In
large pupil; 3) cracks on iris; 4) overly small pupil; 5) spots
order to obtain an equal number of normal irises for comparison,
on iris; 6) rough image; and 7) other types. This means that
150 normal irises were randomly picked, as shown in Table I.
ordinary people may easily notice the largeness of the pupil, but
hardly notice the shape of the pupil and the textural differences
B. Manual Blind Clustering Analysis of the irises. However, once people were given some a priori
Our first experiment carried out a blind clustering analysis of knowledge in that the shape of the pupil infers some sort of
the normal irises. Our goal was to observe if specific geometrical diagnostic information, they may pay more attention and will
differences among the irises may draw people’s attention, and perform better classification, as shown by the next experiment.
226 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 17, NO. 1, JANUARY 2013
TABLE III
MANUAL CLASSIFICATION OF NORMAL AND ABNORMAL IRISES
(AVERAGE PERFORMANCE ON TEN RESPONDENTS)
TABLE IV
MANUAL CLASSIFICATION OF NORMAL AND ABNORMAL IRISES BY
INDIVIDUALS WITH MEDICAL BACKGROUNDS (AVERAGE
PERFORMANCE ON SIX MEDICAL RESPONDENTS)
C. Manual Classification of Normal and Abnormal Irises
The second experiment was carried out after the first exper-
iment. Basic iridology knowledge was provided to the ten se-
lected respondents, and they were asked to pay attention to the
geometrical shape of the pupillary boundary and the largeness
of the pupil. Then, they were asked to complete the following
task: “Given 292 iris images, please determine which one is
normal or abnormal.”
This is a manual classification task. The result is given in
Table III. It can be seen that given some a priori knowledge
about the pathological expression of typical geometric features
of the iris, ordinary people may distinguish abnormal irises with
a relatively good performance.
D. Manual Classification of Pathological Irises by Who
With a Medical Background
Our third experiment involved in total six respondents with
a medical background, three individuals from the eye clinic de-
partment in each of our collaboration hospitals. However, there
was no iridologist because a professionally qualified iridologist
is difficult to find. The second experiment was repeated for these
six medical experts than those ten regular participants who took
part in the second experiment. The result is given in Table IV.
It can be seen that the number of normal irises that are mis-
assigned as abnormal has more than tripled for those with a
medical background in contrast to regular participants. On the
contrast, very few abnormal irises are misassigned as normal
irises. This shows that those who with a medical background
tend to be overly careful and cautious, the reason is perhaps due
to their professional attitudes.
Therefore, from the three experiments that involved manual
analyses of irises, we can conclude that certain geometric fea-
tures of an iris have more or less pathological significance, and
it is possible to apply them to aide clinical diagnoses of some
diseases, such as gastrointestinal ones.
Fig. 3. Schema of pupil deformation, such as reduction and dilatation of the
pupil, and an irregular boundary.
III. GEOMETRIC DEFORMATION FEATURE EXTRACTION
As mentioned previously, iridology cites that there exists a
clear correlation between some geometrical features and some
common and frequently occurring diseases. Through a large
amount of long-term clinical observations, gastrointestinal dis-
eases have been proven to cause some distinct geometrical fea-
tures on the iris, such as those that are related to the form of
the pupil, smoothness of the pupillary boundary and largeness
(the proportion of the pupil that occupies the iris), and similar
geometric deformations in the collarette.
In order to quantitatively describe the geometric deformation
of the iris and provide a method to merge the characteristics of
pupils and collarettes together, we propose here some quantified
measures on geometric features in irises with pathological con-
ditions based on computer-aided image processing techniques.
A. Pupil-Based Features
By observing the samples in our research database, it was
found that individual pupils and irises have great differences.
However, the ratio between the pupil and iris, the roughness
measures of the geometry of the pupil edges, and other geometric
characteristics of the pupils, indicate large differences between
a regular person and patients with various diseases.
We propose to study the geometric characteristics of pupils
based on the following three viewpoints.
1) Pupil Roundness Feature L1 : Pupil roundness is the ratio
between pupil diameters in various directions. Some diseases,
such as the gastric ulcers and gastric cancer will damage the
inner edge of the pupil, and as the result, the pupil shape, espe-
cially its roundness, will change. A typical phenomenon is that
the pupil can no longer maintain a standard round shape, and
becomes oval or even an irregular shape, as shown in Fig. 3.
Pupil roundness feature L1 is defined as the ratio between the
maximum value and the minimum value of the four measure-
ments of the pupil diameter, as shown next:
 
max Dhinner , Dvinner , Dlinner , Drinner
L1 =   (1)
min Dhinner , Dvinner , Dlinner , Drinner
where Dhinner refers to the horizontal diameter of the pupil,
which is measured by the horizontal width of the pupil, Dvinner
is the vertical diameter, and Dlinner and Drinner comprise the
diameters in the directions at 45◦ and 135◦ , respectively.
It can be easily seen that in standard circles L1 = 1.0, and a
noncircular shape will mean that L1 > 1.0. A larger value of L1
means a much irregular shape.
Clearly, the more the number of directions chosen for diam-
eter estimation, the more the possibility that the real maximum
MA et al.: IRIS-BASED MEDICAL ANALYSIS BY GEOMETRIC DEFORMATION FEATURES
TABLE V
MEAN VALUE OF L 1 UNDER SEVERAL CONDITIONS FOR ESTIMATING THE
DIAMETER OF THE PUPIL
TABLE VI
STATISTICAL ANALYSIS OF PUPIL FEATURES FOR NORMAL AND
PATHOLOGICAL IRISES
and minimum diameters of the pupil can be determined; hence,
a larger L1 value is obtained. Table V shows our experimental
results for choosing the best number of directions. We tried to
estimate diameters for six cases: horizontal and vertical direc-
tions, the 4 directions as indicated previously, and 8, 12, 16, and
32 directions that are equally divided in a circle. We calculated
the L1 value for all of the 142 pathological iris examples in
our database, and provided the mean value of L1 in the table.
It is apparent that more directions augment a mean L1 value,
and could make the L1 feature more significant. However, the
computational cost was also greatly augmented. The four direc-
tions are chosen for a good balance between performance and
computational costs.
2) Pupil Largeness Feature L2 : The pupil largeness feature
is the average ratio of the pupil diameter to the iris diameter,
and describes the amount of pupil that occupies an eye. When
disease causes the tension in the inner edge of the iris to change,
either tightening or breaking, the pupil is reduced or enlarged.
As the result, the largeness of relative area of the pupil in the
iris is changed and reflected in the pupil iris ratio.
Here, pupil iris diameter ratio L2 is defined as the average of
the horizontal, vertical, and two directional ratios between the
pupil and the iris diameter
 
1 Dhinner Dvinner Dlinner Drinner
L2 = + + + (2)
4 Dhouter Dvouter Dlouter Drouter
where Dhouter , Dvouter , Dlouter , and Drouter are the iris widths in
the horizontal, vertical, 45◦ , and 135◦ directions.
3) Pupil Edge Roughness L3 : This neatness and smoothness
feature in the inner edge of the iris is very important. It is the
most important indicator of a healthy digestive system. Given
an edge that contains points (x1 , x2 , . . . , xk , . . . , xK , x1 ), the
L3 roughness feature is defined as:
1 
K
L3 = |ρk − ρ̄|. (3)
K
k =1
Here, ρk is the curvature at element xk , and ρ̄ stands for the
mean curvature. In our research, a third-order curvature is ap-
plied and ρk ∈ [0, 2π] is measured as the angle between the line
(x(k + 3) m o d K − xk ) and the line (x(k −3) m o d K − xk ), as what
227
Fig. 4. Distribution of the examples in our database according to the L 1
feature.
(x −x )•(x −x )
is given by ρk = arccos |(x (( kk ++ 33 )) mm oo dd KK −x kk )||(x (( kk −3 ) mod K k
−3 ) m o d K −x k )|
.
Obviously, a greater L3 value indicates increased worsening
in the degree of smoothness.
By using the previous definition of the pathological features
of an iris pupil, we analyzed 150 normal irises and 142 irises in
patients with digestive pathology. The statistic results are given
later. Obviously, the majority of the parameters of normal and
pathological irises have good classification results.
Fig. 4 shows a histogram of the iris examples in our database
based on the L1 feature. The same phenomenon can be seen
for the L3 feature. However, the L2 feature does not have equal
differentiating power, as shown in Table VI.
B. Collarette-Based Features
It was also observed that the area between the pupil boundary
and collarette contains large variations. Normal collarettes form
a jagged closed loop. Diseases may cause collarettes to expand
or reduce with geometric changes. The collarette boundary can
be estimated through an ant colony optimization based image
segmentation algorithm (ACO-ISA) proposed in our former re-
search [17]. Similar to pupil pathological features, collarette
features may be quantitatively defined and calculated for iris
diagnostic applications. The collarette features proposed in this
paper are as follows.
1) Roundness of Collarette Boundary A1 : A1 is defined as
the ratio between the maximum and the minimum values of the
six diameter measurements of a collarette, as shown below:
 
max Dhcollar , Dvcollar , Dlcollar , Drcollar , Dm collar
, Dncollar
A1 =  
min Dhcollar , Dvcollar , Dlcollar , Drcollar , Dm collar , D collar
n
(4)
where Dhcollar , Dvcollar , Dlcollar , Drcollar , Dm
collar
, and Dncollar are
the collarette diameters in the horizontal, vertical, 30◦ , 60◦ ,
120◦ , and 150◦ directions. The reason that these six diameters
are used for collarette features instead of the four that were used
for the pupil is that the collarette boundary is much rougher
than the pupil boundary, and 6 is a good balanced number of
diameters in this case.
Clearly, standard circles have A1 = 1.0; and statistics for
a normal person produce a collarette that has a nearly round
boundary with 1.13 < A1 < 1.27.
2) Collarette Iris Diameter Ratio A2 : The anatomy of an
iris reveals that the collarette makes up 1/3 of the iris area. A2
228 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 17, NO. 1, JANUARY 2013

TABLE VII
STATISTIC RESULTS FOR COLLARETTE FEATURES

is defined as the average diameter ratios in the six directions, as

described by the following formula:

1 Dhcollar Dcollar Dcollar Dcollar Fig. 5. Distribution chart of normal and pathological irises by comparing L 1
A2 = outer + vouter + louter + router and L 3 .
6 Dh Dv Dl Dr

Dcollar Dncollar
+ m outer
+ . (5) 2) Iris and Collarette Boundary Localization: Accurate iris
Dm Dnouter localization is a precondition of feature extraction. The purpose
Here, Douter is the iris diameter in the corresponding of iris localization is to eliminate occlusions, such as those
direction. by eyelids, eyelashes, and pupils, which are useless for iris
3) Collarette Pupil Diameter Ratio A3 : Similarly, A3 is diagnosis. In other words, this is the localization of the inner
given by and outer limbus of irises. Iris localization includes rough and
precise iris localization.

1 Dhcollar Dvcollar Dlcollar Drcollar Localization of iris: In this paper, a partitioning method will
A3 = + + + be used to roughly localize the inner and outer limbus. Our
6 Dhinner Dvinner Dlinner Drinner
collar
 method estimates the gray-level threshold by using a gray-level
Dm Dncollar distribution information of iris images, and then separates the
+ inner + inner . (6)
Dm Dn pupil from the iris image. Finally, the traditional Daugman al-
gorithm is used for rough localization [8].
Fifty normal irises and 37 irises of patients with digestive
Collarette boundary localization is based on the ACO-based
system diseases were analyzed by applying these collarette fea-
iris segmentation method proposed in our former work [17] for
tures. Here, the digestive system diseases were mainly related to
collarette extraction and texture analyses. The idea is to intro-
intestinal system illnesses. The results are shown in Table VII.
duce information about the problem into the decision making
Obviously, the majority of the parameters of normal and patho-
process and the pheromone releasing process of the artificial
logical irises have better classification results.
ants: 1) an ant moves according to the pheromone intensity
and movement momentum. Different criteria of movement mo-
IV. EXPERIMENTAL RESULTS mentum and pheromone evolution may fit various application
A series of experiments were implemented on the proposed problems; 2) The release of pheromone is monotone to the
pathological iris features. Classification between both normal difficulty of the movement. Movement difficulty considers the
and pathological irises and disease recognition experiments specific image features, such as object, background, edge, color,
were considered. texture, etc. For collarette boundary localization, the histogram
distance between local windows is applied as local texture sim-
A. Necessary Iris Image Pre-Processing for Feature ilarity measure. The movement difficulty is defined by the pixel
Extraction gradient and is inverse proportional to pheromone release.

In order to realize automatic feature extraction, the iris images

should undergo preprocessing as follows: light spot correction,
iris localization, and collarette localization.
1) Light Spot Correction: Due to the current acquisition
equipment and the environment during the acquisition in this
study, bright light spots appeared on the iris images which are
uniformly distributed. Gamma correction is an effective way to
overcome such influences by light [18]. Gamma correction is
based on the estimation of a gamma function for each image,
and gives a nonlinear mapping relationship between gamma and
pixel values to weaken the effects of light. A specific value is
created for the pixels in order to ensure an appropriate method
of correction, which will more accurately make corrections to
pixels that are more severely affected by light [19], [20].
B. Classification Between Normal and Pathological Irises
In consideration of the six geometric features of each iris
proposed in this paper, a series of experiments were realized to
evaluate the classification ability of the proposed features and
feature extraction methods between normal irises and those with
diseases.
Figs. 5 and 6 provide a distribution chart that compares
normal and pathological irises by comparing L1 versus L3 ,
and L1 versus L2 .
It can be clearly seen in Fig. 5 that L1 and L3 have good
classification abilities. The classification ability of L2 depends
on the disease. From the human eye anatomy, it can be seen
that the pupil occupies about 1/3 of the iris in a normal person.
MA et al.: IRIS-BASED MEDICAL ANALYSIS BY GEOMETRIC DEFORMATION FEATURES 229

Fig. 6. Distribution chart of normal and pathological irises by comparing L 1

and L 2 .

Fig. 8. Distribution chart of normal and pathological irises by comparing

collarette shapes.

Fig. 7. Iris images from a healthy person and those suffering from stomach
diseases. (a) Gastric ulcer. (b) Gastroduodenal bleeding. (c) Gastric cancer.
(d) Normal.

However, for individuals with stomach illnesses in Fig. 7, the

pupil may shrink for some patients, and may also enlarge for
others. Thus, the average value of L2 poorly contributes to the
classification of digestive diseases.
However in our further research, when inflammation and
bleeding cases are separately considered, it is found that inflam-
mation usually leads to pupil contraction, and on the contrary,
gastroduodenal bleeding and more severe types of illnesses will
reduce the tension of the inner edge of the iris, so that the pupil
is enlarged. Hence, we are able to further distinguish between Fig. 9. Combination of features and performance on disease classification.
these two kinds of gastrointestinal diseases.
Similar analyses on classification ability were completed for
The experiments were designed to identify two typical types
A1 , A2 , and A3 , as shown in Fig. 8. It is observed that these fea-
of diseases: gastrointestinal and enteritis system diseases. In or-
tures also have good classification abilities. However, the num-
der to create a baseline comparison with the SOM and SVM
ber of ambiguous examples is somewhat higher than that for
methods, a classical pattern recognition method—Fisher linear
pupil features. The reason is that collarette contour extraction
classifier—is also implemented for the iris identification exper-
is more difficult and less precise, and leads to a wider distri-
iments, and the results are shown in Fig. 9.
bution of collarette features and influences their classification
In order to determine the classification ability of the pro-
effectiveness.
posed methods, a principal component analysis (PCA) [24] was
conducted on the entire set of features, including the proposed
C. Disease Recognition
pupil- and collarette-based features. A distribution of all the
Since the number of examples in our iris database is consid- iris examples according to the two most important components
ered insufficient, we chose a self-organizing map neural network (noted PCA1 and PCA2) is given in Fig. 10.
(SOM) and support vector machine (SVM) for classifying dis- In the investigation of the performance of the three pat-
eases, which were used in our former research work [21]–[23]. tern recognition methods in our experiments, it is easy to find
230
Fig. 10. Gastrointestinal disease, enteritis system disease, and normal iris
distribution per the first two main features given by a PCA analysis based on
pupil and collarette shapes.
TABLE VIII
CLASSIFICATION PERFORMANCE ON TYPICAL DISEASES VERSUS
HEALTHY PERSON
that the Fisher linear classifier method is simple, but gives a
slightly lower recognition rate; the SVM method may be the
best in recognition, but to take into account more features in the
experiment requires the combination of prior knowledge and
more direction, so its recognition performance is lacking; and
finally, the SOM network method is also very effective, because
the number of neural network units can be further added into
the model to enhance its capacity, and also because no a priori
classification knowledge is needed for the exact classification of
an object. Therefore, we believe that the SOM network method
may be more increasingly competitive in future practical use of
iris diagnostic systems.
Table VIII shows the average recognition rate of the three
methods in this practical disease recognition application. All
three methods achieve nearly a 90% recognition rate for the indi-
viduals with the two diseases and the healthy individuals. These
good recognition rates also prove the clinical diagnosis signifi-
cance and the effectiveness of the proposed pupil and collarette
features. These features allow a large degree of choices for clas-
sifiers and imply great potential for practical applications. A
pair t-test reveals no significant difference (p > 0.05) between
our method and manual classification in terms of recognition
rates, and thus further proves the meeting of the requirements
for clinical applications.
In further comparing the performance of the system with our
former research, it can be seen that for a single internal organ,
the proposed pathological features may produce a high degree
of specificity with a high recognition rate, such as the case of
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 17, NO. 1, JANUARY 2013
nervous system diseases. Since the digestive system includes the
stomach, intestines, liver, gallbladder, and other organs, patho-
logical changes are relatively much more complex, so the iris
features also become more complex, and slightly reduce the
average recognition rate. This means that the proposed clinical
iris feature extraction method is quite promising and may find
further applications in more disease diagnostic applications.
V. CONCLUSION
In this paper, we have proposed a series of geometric features
to quantitatively analyze pathological changes in irises caused
by common illnesses, such as gastrointestinal diseases. A statis-
tical investigation is performed on a database with irises from
normal individuals and patients who suffer from gastrointestinal
diseases, and demonstrated well the diagnostic abilities of the
proposed features. We also proposed a distinctive computerized
iris diagnostic approach, and the experimental results are quite
promising.
The main contribution of this research is that not only is the
mapping relationship between iris characteristics to Western
medicine defined diseases revealed, but also contributes to both
iris diagnosis and recognition by separating the required iris
features into globally structural and locally geometrical features.
This will undoubtedly prompt the modernization of traditional
iris diagnosis, and more importantly, lessen the research gap
between iridology research and clinical applications.
REFERENCES
[1] D. Zhang, Automated Biometrics—Technologies and Systems. Norwell,
MA: Kluwer, 2000, pp. 18–25.
[2] A. Jain, A. Ross, and S. Prabhakar, “An introduction to biometric recog-
nition,” IEEE Trans. Circuit Syst. Video Technol., vol. 14, no. 1, pp. 4–20,
Jan. 2004.
[3] K. W. Bowyer, K. Hollingsworth, and P. J. Flynn, “Image understand-
ing for iris biometrics: A survey,” Comput. Vis. Image Understanding,
vol. 110, pp. 281–307, 2008.
[4] B. Pang, D. Zhang, and K. Wang, “Tongue image analysis for appendicitis
diagnosis,” Inf. Sci., vol. 175, pp. 160–176, 2005.
[5] D. Zhang, L. Zhang, D. Zhang, and Y. Zheng, “Wavelet based analysis of
Doppler ultrasonic wrist-pulse signals,” in Proc. Int. Conf. Biomed. Eng.
and Inf., vol. 2, pp. 539–543, 2008.
[6] D. Guo, D. Zhang, N. Li, and L. Zhang, “A novel breath analysis system
based on electronic olfaction,” IEEE Trans. Biomed. Eng., vol. 57, no. 11,
pp. 2753–2762, Nov. 2010.
[7] C. Oyster, The Human Eye Structure and Function. Sunderland, MA:
Sinauer Associates, 1999.
[8] J. Daugman and C. Downing, “Epigenetic randomness, complexity and
singularity of human iris patterns,” Proc. Roy. Soc. London B, vol. 268,
pp. 1737–1740, 2001.
[9] T. J. Buchanan, C. J. Sutherland, R. J. Strettle, T. J. Terrell, and A. Pewsey,
“An investigation of the relationship between anatomical features in the
iris and systemic disease, with reference to iridology,” Complementary
Therapies Med., vol. 4, pp. 98–102, 1996.
[10] K. Hollingsworth, K. W. Bowyer, and P. J. Flynn, “Pupil dilation degrades
iris biometric performance,” Comput. Vis. Image Understanding, vol. 113,
pp. 150–157, 2009.
[11] J. Daugman, “The importance of being random: Statistical principles of
iris recognition,” Pattern Recognit., vol. 36, pp. 279–291, 2003.
[12] M. J. Hogan, J. A. Alvarado, and J. E. Weddell, Histology of the Human
Eye. Philadelphia, PA: Saunders, 1971, pp. 11–14.
[13] B. Jensen, “L’iridologie, science et pratique de l’art de guerir,” pp. 11–13,
1988.
[14] R. Cong, S. Song, and Y. Liu, “Study of genetic mutation locus in a
family with congenital aniridia,” Chin. J. Ophthalmol., vol. 42, no. 112,
pp. 1113–1117, 2006.
MA et al.: IRIS-BASED MEDICAL ANALYSIS BY GEOMETRIC DEFORMATION FEATURES
[15] R. Roizenblatt, P. Schor, F. Dante, J. Roizenblatt, and R. Belfort, “Iris
recognition as a biometric method after cataract surgery,” Biomed. Eng.
Online, vol. 3, no. 1, pp. 2–8, 2004.
[16] U.S. NLM/NIH Medline Plus, Glaucoma. (2011). [Online]. Available:
http://www.nlm.nih.gov/medlineplus/ency/article/001620.htm
[17] L. MA, K. Wang, and D. Zhang, “An universal texture segmentation and
representation scheme based on ant colony optimization for iris image
processing,” Comput. Math. Appl., vol. 57, pp. 1862–1868, Nov. 2009.
[18] R. P. Kleihorst, R. L. Lagendiik, and J. Biemond, “An adaptive order-
statistic noise filter for gamma corrected image sequences,” IEEE Trans.
Image Process., vol. 6, no. 10, pp. 1442–1446, Jun. 1997.
[19] N. M. Kwok, Q. P. Ha, D. Liu, and G. Fang, “Contrast enhancement and
intensity preservation for gray-level images using multiobjective particle
swarm optimization,” IEEE Trans. Autom. Sci. Eng., vol. 6, no. 1, pp. 145–
155, Jun. 2009.
[20] H. Farid, “Blind inverse gamma correction,” IEEE Trans. Image Process.,
vol. 10, no. 10, pp. 1428–1433, Oct. 2001.
[21] D. Brugger, M. Bogdan, and W. Rosenstiel, “Automatic cluster detection
in Kohonen’s SOM,” IEEE Trans. Neural Netw., vol. 19, no. 3, pp. 442–
459, Mar. 2008.
[22] A. K. Jain, A. Ross, and S. Prabhakar, “An introduction to biometric
recognition,” IEEE Trans. Circuits Syst. Video Technol., vol. 14, no. 1,
pp. 4–20, Jan. 2004.
[23] V. N. Vapnik, S. E. Golowich, and A. J. Smola, “Support vector method
for function approximation, regression estimation, and signal processing,”
Adv. Neural Inf. Process. Syst., vol. 9, pp. 281–287, 1996.
[24] K. R. Tan and S. C. Chen, “Adaptively weighted sub-pattern PCA for face
recognition,” Neurocomputing, vol. 64, pp. 505–511, 2005.
Lin Ma received the M.E. degree from University
Paris 6, Paris, France, in 2001 and the Ph.D. de-
gree from the Harbin Institute of Technology (HIT)
Harbin, China, in 2008, both in computer science.
Since 2002, she has been in the Biocomput-
ing Research Center of Computer Science and
Technology, HIT, where she is currently an Associate
Professor in the Department of Computer Science.
Meanwhile, in 2006 and 2011 she was a Visiting
Scholar in The Hong Kong Polytechnic University
supported by Hong Kong Croucher Funding. Her re-
search interests include intelligent information processing, image processing,
and cognitive science.
David Zhang (F’09) received the Graduate degree in
computer science from Peking University, Beijing,
China, and the M.Sc. degree in computer science
and the Ph.D. degree from the Harbin Institute of
Technology (HIT), Harbin, China, in 1982 and 1985,
respectively. He received the second Ph.D. degree in
electrical and computer engineering from the Univer-
sity of Waterloo, ON, Canada, in 1994.
From 1986 to 1988, he was a Postdoctoral Fellow
at Tsinghua University, Beijing, and then, an Asso-
ciate Professor at the Academia Sinica, Beijing. He
is currently a Head in the Department of Computer, and a Chair Professor at
the Hong Kong Polytechnic University, Kowloon, Hong Kong, where he is
the Founding Director of the Biometrics Technology Centre supported by the
Hong Kong SAR Government in 1998. He was also a Visiting Chair Professor in
Tsinghua University, and an Adjunct Professor in Peking University, Shanghui
Jiao Tong University, HIT, and the University of Waterloo. He is the Founder and
Editor-in-Chief of International Journal of Image and Graphics, Book Editor of
Springer International Series on Biometrics, an Organizer of the International
Conference on Biometrics Authentication, an Associate Editor of more than ten
international journals, including IEEE TRANSACTION AND PATTERN RECOGNI-
TION. He is the author of more than ten books and 200 journal papers.
Dr. Zhang is a Croucher Senior Research Fellow, Distinguished Speaker of
the IEEE Computer Society, and a Fellow of the International Association of
Pattern Recognition.
231
Naimin Li received the Graduate degree from the
Department of Medical Treatment, Traditional Chi-
nese Medicine (TCM), Shenyang, China, in 1962.
He is currently a Fellow Professor at the Harbin
Institute of Technology, Harbin, China. Since 1965,
he has been involved in the study and exploration of
the application of tongue diagnosis in modern dis-
ease diagnosis and treatment, in the fields of internal
medicine, pediatrics, epidemiology, and gynecology.
During 1970 and 1972, he was a Leader of Experts
Group of TCM by the Chinese government in the
Middle-Europe and Albania. In 1989, he started the first tongue image labo-
ratory in China. He has authored or coauthored more than 250 papers and 14
books around his research areas.
Yan Cai received the M.S. degree from Harbin Med-
ical University, China, in 2001 and the Ph.D. degree
from the Tongji Medical School at Huazhong Uni-
versity of Science and Technology in 2012.
She is now a Professor and Chief of Maternity De-
partment at the Fourth Affiliated Hospital of Harbin
Medical University, Harbin, China, and the Deputy
Director in Perinatology Medical Association of
HeiLongjiang Province. She has authored or coau-
thored more than 30 papers and 2 books around her
research areas. She is involved in gynecology and ob-
stetrics clinical work for 24 years, nearly 10 years mainly involved in obstetrics
work. Her research interests include bioinformatics, perinatology, and high-risk
pregnancy.
Prof. Cai received seven provincial R&D awards and one national patent.
She has also received the Stefano Mauritania national level medal.
Wangmeng Zuo (M’09) received the Ph.D. degree
in computer application technology from the Harbin
Institute of Technology, Harbin, China, in 2007.
He is an Associate Professor in the School of
Computer Science and Technology, Harbin Institute
of Technology. From July to December 2004, from
November 2005 to August 2006, and from July 2007
to February 2008, he was a Research Assistant in the
Department of Computing, Hong Kong Polytechnic
University. From August 2009 to February 2010, he
was a Visiting Professor in Microsoft Research Asia.
His research interests include sparse representation, biometrics, pattern recog-
nition, and computer vision.
Dr. Zuo is an Associate Editor of IET Biometrics and the reviewers of several
international journals.
Kuanquan Wang (SM’07) was born in Sichuan
Province, China, in September 1964. He received
the B.E. and M.E. degrees in computer science from
the Harbin Institute of Technology (HIT), Harbin,
China, and the Ph.D. degree in computer science
from Chongqing University, Chongqing city, China,
in 1985, 1988, and 2001, respectively.
From 1988 to 1998, he was with the Department
of Computer Science, Southwest Normal University,
Chongqing, China. Since 1998, he has been with the
Biocomputing Research Center of Computer Science
and Engineering, Department of HIT, where he is currently a Professor at the
Department of Computer Science and Engineering, and an Associate Director
of the Biocomputing Research Center in HIT. Meanwhile, from 2000 to 2001,
he was a Visiting Scholar in The Hong Kong Polytechnic University supported
by Hong Kong Croucher Funding, and from 2003 to 2004 he was a Research
Fellow in the same university. So far, he has published more than 90 papers.
His research interests include biometrics, image processing, pattern recognition,
and biometrics-based diagnosis technology for traditional chinese medicine.
Dr. Wang is an Associate Editor of International Journal of Image and
Graphics, a reviewer of IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBER-
NETICS, Pattern Recognition.