Fundus Foveal Localization Based on Image Relative Subtraction

Document Sample
Fundus Foveal Localization Based on Image Relative Subtraction Powered By Docstoc
					          Fundus Foveal Localization Based on Image
             Relative Subtraction-IReS Approach
                            Jeetinder Singh                                                 Jayanthi Sivaswamy
          International Institute of Information Technology                International Institute of Information Technology
                           Hyderabad,India                                                  Hyderabad,India
                 Email:                              Email:

   Abstract—In this paper we present a method for fovea localiza-
tion which does not use organization information of other retinal
structures like optic disk and arcades. The main advantage of
this method is that it does not require segmentation/localization of
other retinal structures which are required as prior knowledge in
existing fovea localization methods. The key idea is to enhance the
relative contrast between the fovea and its surrounding such that
it is well-separated in a retinal image. The approach was tested
on 520 retinal images which include normal and pathological
color retinal images. An overall accuracy of 90.57% is reported
independent of left eye, right eye, normal or pathological retinal
images.                                                                                  (a)                                 (b)
                       I. I NTRODUCTION                                Fig. 1.   (a) Retinal image with marked fovea rgion (b) Gray channel image
   Fovea is a key region which is responsible for high acu-
ity colour vision. Hence, the location of lesions relative to
fovea, is of interest in diseases such as age-related macular
degeneration or diabetic retinopathy. This information is used         used to locate the fovea. OD is localized using principal com-
by ophthalmologists for correct diagnosis and treatment. In            ponent analysis and modified active shape model is applied
general, estimation of the fovea shape and localization helps in       to extract the main courses of blood vessels. Localization of
establishing statistics regarding position of lesions for disease      the fovea is done based on its geometrical relation to OD and
gradation and provide a better description of other features           blood vessels. The approach showed a sensitivity of 100% on
in fundus images. Accordingly, fovea detection forms a com-            35 images. A similar strategy [3] was employed based on first
ponent of most systems that are designed for automatically             identifying the main blood vessels using the modified active
analysing retinal images. The challenge in fovea detection is          shape model. The candidate region is determine by spatial
that its size and/or position can vary with the chosen field            relationship of the OD and the parabolic structure of blood
of view (macula centric or optic disk centric), magnification           vessels. The darkest region is searched within candidate region
level, non-uniform illumination; the fovea can also suffer from        to identify fovea.
partial or full occlusion due to pathological factors such as             In [4] approach is design based on a priori knowledge of
lesions, scars etc. The existing approaches employed for fovea         OD position and its relative distance from fovea. Template was
localisation use geometrical relation to other structures to           defined to approximate a typical fovea. The fovea is identified
locate the fovea.                                                      based on the location of maximum correlation between the
   A method [1] is designed based on priori knowledge of               template and the intensity image in HSI space. The variance
the optic disk (OD) position . The OD is localized based               of intensity of adjacent pixels was used to locate OD. The
on pyramidal decomposition and Hausdorff-distance based                method was tested on a total of 112 retinal images divided
template matching. The candidate regions for OD are first               into 3 groups and detected fovea in 78 images.
found based on pyramidal decomposition on the grayscale                   In [5] , a system closely related to [2] is designed to
green plane of the original image. Candidate regions are               determine the landmark positions by fitting a single model to
looked for the presence of circular templates using Hausdorff-         the image. The algorithm requires two sub models. One is a
distance as a distance measure. The OD is identified based              shape model and second model is an energy function. A set of
on center of the fitted circular template. Macula distance with         16 points were chosen to define shape model known as a point
respect to the OD is used as a priori knowledge to position            distribution model. The algorithm finds the shape or landmarks
fovea. The method reported 95% accuracy on 40 images.                  that minimize energy function. A number of different system
   A model based approach is proposed [2] in which the OD              setups are compared. The approach reported 94.4% accuracy
location and extraction of the main courses of blood vessels is        on the 500 images of the screening test set and 92.0% on the
                                                                   to be the required plane Ia as it contains the fovea (which
                                                                   appears as a dark region) embedded in clutter made up of
                                                                   vessels (which also appear dark), optic disk (bright region)
                                                                   and other pathologies. These characteristics can be observed
                                                                   in the gray level intensity plane of image in Fig.1,(b). It can be
                                                                   observed from this image that the contrast of the foveal region
         (a)                     (b)                    (c)        is not very high as the overall appearance of the intensity
         Fig. 2.   Few samples fundus images with pathology        plane is on the dark side. Alternatively, the signal of interest
                                                                   is biased with a low intensity background illumination which
                                                                   itself can be slowly varying across the image. This observation
                                                                   serves to guide the choice for the image plane Ib . A suitable
100 pathological test images.                                      choice for Ib should contain the foveal region in addition
   The goal of this paper is first develop a method for fovea       to providing information about the background illumination.
localization without prior knowledge of other features. Sec-       The red plane image which is shown in Fig.4(a), is a good
ondly, the algorithm is analyzed and tested for detection of       candidate for Ib . The red plane’s visual appearance indicates
fovea region. The approximate size of the visible fovea region     that the dominant component is the illumination function
in a final output image can also be obtained.                       (bright and slowly varying). This point has been exploited to
   We present an algorithm for detecting the fovea which is        perform illumination correction, via histogram matching with
based on the following domain knowledge: fovea is the darkest      the green channel image, prior to vessel detection in [6]. Other
region within the macula in a colour retinal image with a          attractive features of the red plane for our problem is that
distinguishing feature namely, the absence of blood vessels        it contains the fovea as a low contrast region in addition to
(Fig.1). The degree of darkness and size depends on factors        another component of the clutter, namely, the vessel structure
such as uneven illumination, presence of lesions such as drusen    (which share similar characteristic as the fovea in the image
and exudate’s. The proposed algorithm is based on difference       Ia ). Hence, by combining the two gray level intensity and red
of two image planes by modeling the fovea detection as one         plane images it should be possible to extract the foveal region.
of signal detection from clutter. The outcome of the proposed      The strategy for combining should ensure that the relative
detection consists of top two candidates for fovea. An effort      contrast of the foveal region is strengthened while that of other
has been made to cope with retinal images distorted by the         structures are weakened. The proposed algorithm is based on
presence of lesions etc.                                           difference of two image planes by modeling the ROI as a dark
   The paper begin with brief discussion devised a very            region in given input image and defined as:
basic definition of relative subtraction to describe the fovea
localization algorithm in section II. In section III, we discuss                           Is = Ia − (Ib )c                      (1)
fovea localization approach. And finally will conclude with
experiment results and a brief discussion.                            Where Is is Ia ∩ Ib , set which contain common element of
                                                                   Ia and Ib .
II. R EGION OF I NTEREST (ROI) AGGREGATION BASED              ON      We will now show the operations that are required to achieve
         I MAGE R ELATIVE S UBTRACTION - IR E S                    this goal.
   The Fig.1,(a) shows a sample colour image taken with               1) Selection of Ia and Ib : In image Ia the (ROI) is em-
a macula-centric field of view. It can be observed that the                bedded. The acquisition of image Ib should be designed
fovea is a dark and roughly homogeneous region. However,                  such that it contains the (ROI) in addition stretch the
the relative contrast between the fovea and its surround can              brighter region over wider range without distorting the
decrease with an increase in the magnification level. This can             ROI characteristic.
be seen from fig.2. From a signal detection perspective, fovea         2) Calculate Is = Ia − (Ib )c
detection can be viewed as a problem of detecting the signal of       3) Apply threshold T as defined distance measure for fil-
interest (fovea) from a given colour image where the signal is            tering plausible candidates for ROI .
embedded in clutter. The clutter component represents all the         4) Apply prior knowledge to correctly identify the object
background structures as well as the non-uniform illumination             ROI.
in the image. Let us consider projecting the image into two           Theoretically step 2 will enhance common regions in image
appropriate planes Ia and Ib such that both Ia and Ib contain      Ia and Ib . If ROI, in our case fovea, share the same charac-
the signal of interest. Additionally, let Ib capture the clutter   teristics in all three grayscale image channels , the selection
to a large extent. Subtracting Ib from Ia will remove the          of two grayscale image channels will ensure the enhancement
clutter from Ia but will also weaken the signal. If instead        of ROI. The approach is simple and works if object Ir can be
we complement Ib and do a relative subtraction operation, the      associated with unique characteristic to finally filter it from the
signal gets enhanced while the clutter is suppressed.              result of subtraction. Fig.3 illustrates the flowchart of proposed
   In the context of fovea detection, the plane that contains      algorithm. Illumination corrections and contrast enhancement
achromatic input, namely the intensity plane can be taken          are required in the presence of variety of retinal diseases.
                        Fig. 3.   IReS algorithm

          (a)                      (b)                     (c)
Fig. 4. Three component of color image: (a) Red Channel (b) Green Channel
(c) Blue Channel                                                                       (a)                      (b)                      (c)
                                                                            Fig. 5. Each row shows result of selection of three different Ib images
                                                                            expressing properties of Eq.1, where Ia in each row is gray channel image.
                                                                            The images Ib in each row are: red channel, green channel, blue channel from
                                                                            top row to bottom. The gray scale distribution of image Ib in each row was
     III. F OVEA L OCALIZATION BASED ON R ELATIVE                           described by combination of local adaptive contrast enhancement and linear
                S UBTRACTION A PPROACH                                      contrast stretching function window (low-in,high-in,0,1).

A. Selection of two images

   The proper selection for image Ia and Ib is essential
such that in later stages we obtained a enhanced grayscale
image in which ROI gets maximum contrast compare to
other retinal structures. In Fig.5 experiment analysis of Eq.1
clearly illustrates the effect of choosing the different Ib . The
image Ia is chosen as contrast enhanced gray channel in
RGB space where the (ROI) is embedded. Theoretically the                                       (a)                               (b)
image Ib image should contain the (ROI) as dark region as
foreground and rest of the retinal image as bright background.              Fig. 6. (a) Enhanced resultant gray scale image. (b) Contrast enhancement
                                                                            of the the image shown in Fig.6 (a) using iterative contrast enhancement
The main idea behind the approach is that the image Ib                      approach.
should contain only the (ROI) foveal region in addition to the
background illumination (Fig.5). The red channel image is nor-
mally brighter and captures the maximum noise in the retinal
image, (Fig.4(a)). Initial evaluation showed that approximation             B. Relative Subtraction of Images
of contrast enhanced red channel image as Ib in RGB space                     In Fig.5 experiment analysis of Eq.1 clearly illustrates
perform better in comparison with other color plane images.                 that after subtraction of two images low value pixels are
The uneven illumination and certain abnormalities effect the                enhanced and get distinguished irrespective of its background
contrast between fovea and background. Thus, in order to                    and global threshold is capable of separating the ROI from its
compensate for uneven illumination, contrast enhancement are                background. But in the retinal images there are other dark
required on the gray channel and the red channel images. The                region and as a result selection criteria can easily fail in
gray channel image is darker,(Fig.1(b)), and fovea is not well              pathology effected images. Next to overcome this challenge
separated from its background. The preprocessing is required                gradient method are evaluated.
to separate the fovea from its background without effecting                   Gradient based methods can easily enhance heterogeneity in
the fovea region. Hence a linear contrast-stretching function               the retinal image due to the vessels. Other regions in retinal
is applied to separate the fovea from its background. But the               image can be looked as non homogeneous region in the sense
red channel image of a retinal image tends to be saturated and              that blood vessels are flowing through these regions. Enhance-
visually bright. To enhance the details of local darker regions             ment of non-uniformity will separate the non homogeneous
local adaptive contrast (CLAHE) is applied.                                 and homogeneous regions. But gradient based method will
                                                                             without affecting the fovea region. A sensitive algorithm needs
                                                                             to be designed to identifying the fovea in the presence of
                                                                             heavily confounding features. The top-hat has advantage over
                                                                             complement in the presence of uneven noise and dark regions
                                                                             at periphery. The fig.8 shows schematic representation of
                                                                             modified and improved IReS algorithm for fovea localization.
                                                                                        IV. E XPERIMENT R ESULTS      AND   D ISCUSSION
Fig. 7.   Two-stage contrast enhancement and global thresholding procedure
                                                                                The evaluation of the proposed algorithm was performed
                                                                             using the public database DRIVE [8] , STARE [9] and LVPEI
                                                                             [10]. The images in LVPEI dataset were acquired using a Zeiss
reduce the contrast between between fovea and background                     digital fundus camera with a 50 degree field of view (FOV).
when compared to complement operation. The subtraction                       Each image was captured using 8 bits per color plane at 768
after top-hat morphological operation on red channel image                   by 576 pixels.
(Fig.6(a)) shows how non homogeneity has been brought                           The total number of images taken from each dataset are
into image using top-hat operation near periphery. The top-                  given in table table I. Each database was divided into two
hat has advantage over others as it exploits the circular nature             groups, level-1 and level-2 irrespective of macula centric or
of fovea and feature of ROI is preserved without distortion.                 non macula centric. Level-1 images have a lot of variation in
Also as a result of subtraction of images the resultant enhanced             contrast, noise, fovea size and illumination. Level-2 are images
gray-level image (Fig.6(a)) has well-separated smooth regions                affected by drusen, fovea region lie near to retinal boundary,
sharing some common trait which can be clustered together.                   size of fovea is small compared to other darker region with
The contrast between regions need to be improved to separate                 variation in illumination and contrasts. Images without fovea
the regions.                                                                 region or almost negligible fovea were not considered.
   The adaptive histogram-based segmentation technique en-                      We report the success of our method to localize the fovea
hances cohesive regions and segment areas that slightly differ               using modified IReS approach alone, without prior knowledge
from their background regions. CLAHE is more effective and                   and relationship of other structure in retinal image. The
works well for both global and local variations (Fig.6(b)). An               method has been evaluated on total of 520 images taken from
iterative local contrast enhancement technique performs very                 DRIVE, STARE and LVPEI. The overall accuracy for the
well in cases of large intensity variations. A similar approach              fovea detection is 90.74%, I. The results shows overall fovea
has been taken in [7]. A two-stage contrast enhancement                      detection accuracy of 92.74% on Level-1 dataset and detection
(Fig.7) method is developed which is more effective than                     accuracy of 88.37% on Level-2 dataset. The results have been
iterative local contrast enhancement technique or single pass                visually verified by physician (few results images are shown
contrast enhancement. The single pass stage applies the global               in fig 9).
threshold to all images but it fails to reconstruct fovea region                The method fails in cases: 1)if fovea region fails to satisfy
in case of extreme intensity variation. But iterative pass stage             maximum sized dark region selection constraint 2) cases where
tries to enhance and reconstruct fovea where single pass                     the red channel is highly saturated and fails to capture fovea
stage fails. The histogram equalization (HE) and CLAHE are                   region . The optic disk centric, high saturated red channel
main method of enhancement. Although they are simple to                      images in STARE database lower the method performance
implement and unsupervised this can result in some minor                     ((Fig.10).
inaccuracies, particularly if any other region is enlarged com-
pared to ROI.                                                                 Category          Level-1            Level-2             Total
                                                                                            Fovea Detected     Fovea Detected        Accuracy
C. Detecting the Best Candidate for Fovea Region                               DRIVE         100% (35/35)             -            100% (35/35)
                                                                               STARE        79.48% (31/39)     87.80% (36/41)     83.75% (67/80)
   As a result of improvement in the contrast due iterative local              LVPEI       94.14% (177/188)   88.47% (192/217)   91.11% (369/405)
enhancement technique fovea region also get enhanced which
                                                                                Total      92.74% (243/262)   88.37% (288/258)   90.57% (471/520)
helps in choosing a global constant threshold. The pixels above
a normalized intensity of 80% in gray scale intensity image are                                        TABLE I
selected. Even after binarization the main obstacle is removal                     P ERFORMANCE EVALUATION OF THE FOVEA LOCALISATION .
of other plausible candidates in extreme intensity variation and
the presence of retinal blood vessels. Since vessels are like
long threads, this property is used to distinguish between the                                        V. C ONCLUSION
fovea from vessels. Finally the region with maximum area                        A new approach has been proposed to extract fovea region
is marked as plausible candidate with maximum confidence                      automatically and robustly in color fundus images without
level.                                                                       prior knowledge of other landmarks. The approach is useful for
   The main difficulty in applying complement is removing                     independent retinal image processing tools. The approximation
the influence of blood vessels and retinal image boundary                     of visible fovea region can provide better feature description.
Fig. 8. Schematic representation of modified IReS approach for fovea region detection algorithm, Top-hat operator exploit the heterogeneity in the retinal
image due to the vessels.

                 (a)                           (b)                              (c)                              (d)                            (e)
                                     Fig. 9.   Few samples of successful results on level-1 and level-2 dataset images.

                                                                                      anatomical landmarks as well.
                                                                                                                       R EFERENCES
                                                                                       [1] L. Gagnon, M. Lalonde, M. Beaulieu, and M. Boucher, “Procedure to
                                                                                           detect anatomical structures in optical fundus images,” in Proc. of SPIE,
                                                                                           2001, pp. 1218 – 1225.
                                                                                       [2] H. Li and O. Chutatape, “Automated feature extraction in color retinal
                                                                                           images by a model based approach,” in IEEE Trans. Biomed. Eng.,
                                                                                           vol. 51, 2004, pp. 246 – 254.
                                                                                       [3] O. Chutatape, “Fundus foveal localization based on vessel model,” in
                                                                                           IEEE Conf. on Engineering in Medicine and Biology Society, vol. 23,
                                                                                           2006, pp. 1189 – 1195.
           (a)                      (b)                       (c)                      [4] C. Sinthanayothin, J. F. Boyce, H. L. Cook, and T. H. Williamson,
                                                                                           “Automated localisation of the optic disc, fovea, and retinal blood vessels
Fig. 10. Failure Results: (a) Original image (b) Corresponding enhanced red                from digital colour fundus images,” in Br. J. Ophthalmol, vol. 83, 1999,
channel image (c) Gray-scale intensity image after subtraction of two images.              pp. 902 – 910.
                                                                                       [5] M. Niemeijer, M. D. Abrmoff, and B. van Ginneken, “Segmentation
                                                                                           of the optic disc, macula and vascular arch in fundus photographs,” in
                                                                                           IEEE Trans. on Medical Imaging, vol. 26, 2007, pp. 116 – 127.
Also the prerequisites of detecting main vessel branches and                           [6] N. M. Salem and A. K. Nandi, “Novel and adaptive contribution of the
OD is avoided for foveal coordinate system establishment. A                                red channel in pre-processing of colour fundus images,” in Journal of
                                                                                           the Franklin Institute, 2007, p. 243256.
new strategy can be described based on identifying the fovea                           [7] K. Rapantzikos, M. Zervakis, and K. Balas, “Detection and segmentation
region first then OD and finally fitting the parabolic shape to                               of drusen deposits on human retina: Potential in the diagnosis of age-
main vessels.                                                                              related macular degeneration,” in Proc. International Conference on
                                                                                           Image Processing, 2001, p. 10551058.
   The results show that modified IReS method is able localize                          [8] J. Staal, M. N. M.D. Abramoff, M. Viergever, and B. van Ginneken,
fovea region on 520 image with over all 90.57% detection                                   “Ridge based vessel segmentation in color images of the retina,” in
accuracy independent of whether it is dealing with normal or                               IEEE Trans. on Medical Imaging, vol. 23, 2004, pp. 501–509.
                                                                                       [9] “STARE: Structured Analysis of the Retina, .” [Online]. Available:
pathological images. The method shows detection accuracy of                       ahoover/stare/
93.6% on level-1 images and 88.3% on level-2 images. The                              [10] “LVPEI: L V Prasad Eye Institute.” [Online]. Available:
results on pathological images show that the localization of                     
fovea can be used as seed point for the detection of other

Shared By:
Description: Fundus Foveal Localization Based on Image Relative Subtraction