10-08-2012, 02:36 PM
Morphological Image Processing and Network Analysis of Cornea Endothelial Cell Images
Morphological Image Processing.pdf (Size: 511.4 KB / Downloads: 98)
Abstract
In this paper, we propose a robust and accurate method for segmenting grayscale images of corneal endothelial
tissue. Its rst step consists in the extraction of markers of the corneal cells using a dome extractor based on
morphological grayscale reconstruction. Then, marker-driven watershed segmentation yields binary images of the
corneal cell network. From these images, we derive histograms of the cell sizes and number of neighbors, which
provide quantitative information about the condition of the cornea. We also construct the neighborhood graph of
the corneal cells, whose granulometric analysis yields information on the distribution of cells with large number of
neighbors in the tissue. Lastly, these results help us propose a model for the corneal cell death phenomenon. The
numerical simulation of this model exhibits a very good match with our experimental results. This model not only
allows us to rene our understanding of the phenomenon: combined with our results, it enables the estimation of
the percentage of cells having died in a given corneal endothelial tissue.
Introduction
The human cornea is a 500 microns thick, transparent tissue which covers the anterior surface of the eye. The cornea
and the ocular lens are the refracting elements of the eye. At the posterior side of the cornea is situated a single,
connected layer of endothelial cells. These cells, polygonal in shape, are about 20 microns across and about four
microns thick. At birth, the corneal endothelial cells are all hexagons. Cells die as the cornea ages, but there is no
new cell division when a cell die: instead, the surrounding cells elongate to ll in the gap and therefore, the average
size of the endothelial cells increases throughout life. The corneal endothelial cells are responsible for the maintenance
of corneal transparency; active transport enzyme systems pump ions into the regions between the cells, the resulting
osmotic force drives
uid out of the cornea into the aqueous humour. In the normal eye, the
uid transport is
in balance with the
uid transport leakage from the aqueous humour into the cornea and corneal transparency is
maintained. In diseased or injured corneas, this balance is disturbed and the cornea swells and loses its transparency.
The analysis of corneal endothelial cell size and shape is extremely important for diagnostic purposes.
In this paper, we propose a new and robust method to segment grayscale images of corneal endothelial cell
tissue. This method is based on advanced morphological transformations [18, 19], including grayscale reconstruction
[25] and watersheds [2, 27]. It allows us to precisely and robustly extract the cell outlines, which is the rst step
towards the automatic analysis of corneal tissues. This method is described in the section 4.
In section 5, we rst derive shape and size measurements from the obtained binary images. Histograms of
cell sizes and number of edges turn out to convey a lot of information on the corneal condition. We then build
neighborhood graphs on the corneal tissue and use morphological granuometries of these graphs [13, 21] to study the
distribution of large and small cells within the tissue.
Lastly, in section 6, we model the cell death phenomenon. Numerical simulations of this model exhibit a very
good match with the results found experimentally in section 5. Comparing the measurements on a given corneal
tissue with the model allows us to estimate the percentage of cells having died, and therefore provide a good estimate
of the condition of the cornea.
Previous work
Several previous studies have reported attempts to use computer assisted image analysis in order to extract cell size
and shape from photomicrographs of the human cornea endothelium. The main problem has been the location of
the cell boundaries in the gray level image obtained from a photograph or a digitized video frame.
The initial work on the problem of edge detection is represented by the works of Laing et al. [10] and Fabian et
al. [4]: photographs or video images were digitized and histogram manipulation followed by thresholding was used for
edge detection of the cell boudaries. More elegant image processing techniques were recently proposed by Cazuguel
et al. [3]: histogram equalization, top-hat lter followed by skeleton by in
uence zones [11], and Yu et al.: adaptive
local enhancement, low pass lter and matched lters with kernels corresponding to dierent orientations. . .
Still, none of these solutions seem to be robust enough to deal with a vast variety of initial images (contrast,
noise level, lighting conditions may greatly vary from one case to the other). Furthermore, there remains a lot of
work to be done on the interpretation of the results (i.e., of the binary images obtained after segmentation). To our
knowledge, no published work has ever gone beyond the simple measurement of cell areas and number of neighbors.
Data Acquisition
In our study, a corneal wide-eld specular microscope [9, 17, 14] was used to obtain the gray level images of a eld
of several hundred cells from normal human patients. The principle of the microscope is that a slit of light is imaged
on the cornea and the region of the interface between the endothelial cell layer and the aqueous humour exhibits a
high re
ecticity due to the shape change in refractive index. The microscope operated in the specular mode: the
specular re
ection of light from the boundary between the endothelial cells and the aqueous humour forms the image.
The cell borders have stronger specular re
ection than the other regions of the cells and this generates the image
contrast. The scattered light from this interface is imaged onto a lm plane in a camera. A
ash of light is used to
eliminate motion eects on the image. The design of the specular microscope eliminated the bright surface re
ection
from the air-cornea surface, since an applanating microscope objective is used to match the refractive index to the
cornea. An example of the images of our test suite is shown in Fig.
Segmentation
We illustrate our segmentation technique on a smaller image, shown in Fig. 2(a). Notice the very irregular lighting,
which makes it impossible to use any direct thresholding mechanism. On this image and in Fig. 1, one can also
observe that the cells can vary quite a bit in size. This remark becomes even more true with damaged or diseased
cornea. Nevertheless, the cells form a contiguous partition of the space: the separation between any two neighboring
cell is a thin dark line of relatively constant thickness. Therefore, it seems natural to try to detect these dark cell
outlines using the top-hat transformation.