Automatic segmentation of contours of corneal cells

https://doi.org/10.1016/S0010-4825(99)00010-4Get rights and content

Abstract

A fully automatic computerized method for segmenting contours of corneal endothelial cells is proposed. As part of the method, scale-space filtering (i.e. Gaussian filtering) is used to achieve tasks different from noise removal. This type of filtering is applied making use of the separability property of Gaussian kernels, avoiding the erosion of images. A variant of unsharp masking is used to considerably increase the visibility of dark areas of images. It is shown how the overflow that occurs when two images are subtracted can be handled to produce better results than normal unsharp masking. The method is exemplified with a low quality specular micrograph. To test the performance of the method, its output is used to automatically calculate the average cell size of images of different samples of tissue and different visual quality. The obtained results are successfully compared to those obtained with a manual semi-automatic method. A method for reading the segmented contours is suggested as well as two shape representations to achieve morphometric analysis of individual cells.

Introduction

Analyses of images of corneal endothelium are routinely done by ophthalmologists for different purposes. This type of analysis is very useful to determine the health status of the cornea in vivo for intact eyes [1], [2], [3], [4]; moreover it is used to evaluate the healing process after transplantation [5], [6]. These analyses are also useful to evaluate the quality of corneal tissue for transplantation [7], [8], [9], [10]. It is worth mentioning that every year thousands of keratoplasties are achieved in the United States. It has also been of interest to analyze the effect of contact lens wear on the cornea [11]. In addition, some researchers do analyze corneal tissue for other purposes [12].

In all those cases mentioned above, the information contained in the contours of the corneal cells is required; so that, it is desirable to isolate the cell contours from the rest of the corresponding image. This process of separating structures of interest from other information on images is called segmentation.

Early segmentation methods of cell contours were fully manual and consisted on tracing the cell contours on a transparent material that were superimposed on photographs of the samples of tissue [12], [13]. Also, back-projection on a screen and tracing has been used [14], [15]. After manual segmentation, a planimeter has been used to measure cell areas [2], [16]. However, manual methods, in general, imply considerable human error and subjectivity; and are very time consuming. A feasible alternative to eliminate human error and subjectivity is to make use of a computer to achieve automatic image segmentation.

Semi-automatic methods, on the other hand, normally make use of a digitizer [2], [16], [17], [18]. The digitizer records the x–y coordinates of the cell apices when a human operator touches them with an electronic pen. This approach is faster than the manual procedures (can take up to 16 min for analyzing 100 cells [17]), but it is also affected by human error and subjectivity.

Due to the need of an automatic method to achieve segmentation of cell contours, several researchers have proposed interesting solutions. One of the first methods for segmenting cell contours that claimed to be automatic was reported by Lester et al. [19]. However, this method required the intervention of an operator to indicate the starting point of each cell contour. In addition, it also required good quality images. Nishi [20], [21] reported a method that was almost automatic. This method required human interaction and it worked fine only when the best of a series of images was selected by an operator. Corkidi et al. [4] reported an ‘automated’ method that also required a selection of images before doing the segmentation, as well as some interaction with an operator during the segmentation process. Another ‘automatic’ method for analyzing images of donor corneas, acquired with an inverted phase-contrast microscope, was reported by Barisani et al. [10]. This method also required interactive picture enhancement in order to produce good segmentation results. In addition, the corneal tissue should receive special treatment before being imaged.

An alternative method for analyzing corneal endothelium was proposed by Masters [22]. With this method the global properties of the cells are characterized with the Fourier transform of the cell boundaries. In this case segmentation was achieved by manually tracing the cell contours. A further development of this technique that does not require segmentation was reported recently [23].

As can be seen, satisfactory full automatic segmentation of corneal cell contours is hard to achieve with standard image processing techniques. Vecchi et al. [24] arrived to this conclusion after comparing four methods for evaluating corneal endothelial cell density. However, in this work a fully automated segmentation procedure is proposed. The performance of the proposed method was compared with that of a semi-automatic procedure, which is currently in use. With this method not only subjectivity and human error are eliminated, but the time required to achieve segmentation of cell contours can be considerably reduced.

Section snippets

Two-dimensional scale-space filtering

In most applications, the purpose of scale-space filtering is to discard structures that are smaller than the standard deviation of the Gaussian filter in use. However, scale-space filtering (Gaussian filtering) can be used for other purposes. In this work, it was used to (1) eliminate the very low spatial frequency component that often appears in images of corneal cells and (2) to close small gaps between structures and recover contours information.

To evolve a two-dimensional signal f(x,y) in

Materials and methods

The raw data used to test our method were 8-bit gray-level images of 512×480. These images were acquired by digitizing black and white photographs with the frame grabber of a system (BioOptics, Inc, Arlington, MA) installed at the Mayo Clinic Ophthalmology Department. The photographs were previously taken with a 35 mm SRL camera attached to a Keeler–Konan specular microscope [25].

The system mentioned above, designed by Dr. Ron Laing [26], is a semi-automatic system for analyzing corneal tissue.

Removal of the lowest frequency

In order to remove the lowest frequency change of intensity, which is clearly apparent in Fig. 1, the original image was scale-space filtered using a Gaussian filter with standard deviation σ=5 (t=25). For the level of noise that is typically found in this kind of images, σ=5 rendered the best results. The smoothed image (shown in Fig. 2) was subsequently subtracted from the original one, yielding the image shown in Fig. 3. It is worth mentioning that at this point the data type used in our

Results

For the sake of clarity, in the previous section some results have already been described. As shown, even though the image chosen for describing our procedure was of considerable low quality, acceptable results were obtained.

The preservation of the size and shape of the original structures is perhaps the most important aspect of a segmentation procedure. In order to show the precision with which the contours of the cells have been segmented, in Fig. 9 the original image appears overlapped,

Discussion

With regard to the final number of complete contours that can be segmented with our procedure, it appears to be good enough to achieve a useful analysis for all the test images. However, the better the quality of the test image the better the results of our segmentation procedure. As shown in Fig. 11, with an original image of better quality, less spurious contours are obtained and a larger number of useful ones are produced.

If we compare the performance of our algorithm with the performance of

Summary

To evaluate tissue for transplantation or a healing process, ophthalmologists routinely achieve morphometric analyses of corneal endothelium. These analyses are based on the contours of corneal cells. In order to eliminate human subjectivity and error, computerized procedures are preferred to achieve this kind of tasks. In this work a fully automatic computerized segmentation method is proposed. The input of this method is a digitized standard specular micrograph and its output is a binary

Acknowledgements

The author wishes to thank Dr. W.M. Bourne, researcher of the Department of Ophthalmology of the Mayo Clinic, for sharing his knowledge and to Dr. Cedric F. Walker, Ex-chair of the Department of Biomedical Engineering of Tulane University, for his support.

Francisco Javier Sánchez-Marı́n was born in Mexico City in 1952. In 1977, he received the BS degree in Electrical Engineering from the Universidad Nacional Autónoma de Mexico. In 1989 he was awarded the ‘Medalla al mérito universitario’, for having obtained the best grades of his class, when he received the BS degree in Experimental Biology from the Universidad Autónoma Metropolitana (Mexico City). In 1991, he obtained the MS degree in Applied Computer Science from the Colegio de Postgraduados

References (39)

  • G.J. Corkidi, R. Marquez, R. Usisima, J. Toledo, E. Graue, Automated in vivo and on-line morphometry of human corneal...
  • G.N. Rao et al.

    Morphological appearance of the healing corneal endothelium

    Arch. Ophthalmol.

    (1978)
  • F.B. Hoefle et al.

    Human corneal donor material: a method for examination before keratoplasty

    Arch. Ophthalmol.

    (1970)
  • W.M. Bourne

    Examination and photography of donor corneal endothelium

    Arch. Ophthalmol.

    (1976)
  • M. Matsuda et al.

    Specular microscopic evaluation of donor corneal endothelium

    Arch. Ophthalmol.

    (1986)
  • A. Barisani et al.

    Automated digital image analysis of organ culture preserved donor corneas

    Ophthalmic Res.

    (1993)
  • B.A. Holden, D.F. Sweeney, A. Vannas, K.T. Nilsson, N. Efron, Effects of long-term extended contact lens wear on the...
  • T. Suda

    Mosaic pattern changes in human corneal endothelium with age

    Jpn. J. Ophthalmol.

    (1984)
  • N. Gullapally et al.

    Cell size-shape in corneal endothelium

    Invest. Ophtalmol. Vis. Sci.

    (1982)
  • Cited by (38)

    • A Review On digital image processing techniques for in-Vivo confocal images of the cornea

      2021, Medical Image Analysis
      Citation Excerpt :

      A number of papers have addressed the problem of correcting the non-uniform luminosity and contrast Nadachi and Nunokawa (1992); Corkidi et al. (1993); Sanchez-Marin (1999); Scarpa et al. (2007); Khan et al. (2007); Ruggeri et al. (2010); Piórkowski and Gronkowska-Serafin (2011); Piorkowski et al. (2015); Fabijańska (2017); Kolluru et al. (2019); Fabijańska (2019); Vigueras-Guillén et al. (2019c); Auksorius et al. (2020). Techniques include a variant of the unsharp masking technique Sanchez-Marin (1999), intensity adjustment by a sinusoidal transformation curve Scarpa et al. (2007), background subtraction Piorkowski et al. (2015); Fabijańska (2017, 2019) and a parabolic correction in Ruggeri et al. (2010). In Piorkowski et al. (2015) histogram equalization is also applied to uniformly distribute intensity levels.

    • CNN-watershed: A watershed transform with predicted markers for corneal endothelium image segmentation

      2021, Biomedical Signal Processing and Control
      Citation Excerpt :

      The earliest corneal endothelium image segmentation methods take advantage of intensity differences between cell borders and cell bodies to perform adaptive thresholding. Then follows some morphological postprocessing aimed at outlier removal, border cleaning, and thinning [26,33,3]. However, applied to specular or confocal microscopy images, thresholding-based methods yield low accuracy due to high sensitivity to uneven tissue illumination and low sensitivity to weak cell edges.

    • Influence of applied corneal endothelium image segmentation techniques on the clinical parameters

      2017, Computerized Medical Imaging and Graphics
      Citation Excerpt :

      Next, the thinning in the third stage assures that cell's sides have the shortest length, thus it is possible to find the parameters for each cell. The method presented in Sanchez-Marin (1999) used unsharp masking and gap closing in the first stage, thresholding in the second, and skeletonisation with pruning in the third. In contrast, in order to create an accurate binarisation method, a Bayesian framework using the assumption of cell border hexagonality was created for the precise location of cell borders whose posteriori probability function was optimised with a simulated annealing approach in Foracchia and Ruggeri (2003).

    View all citing articles on Scopus

    Francisco Javier Sánchez-Marı́n was born in Mexico City in 1952. In 1977, he received the BS degree in Electrical Engineering from the Universidad Nacional Autónoma de Mexico. In 1989 he was awarded the ‘Medalla al mérito universitario’, for having obtained the best grades of his class, when he received the BS degree in Experimental Biology from the Universidad Autónoma Metropolitana (Mexico City). In 1991, he obtained the MS degree in Applied Computer Science from the Colegio de Postgraduados (Estado de México). From the fall of 1991 to the winter of 1995 he completed the program for obtaining the PhD degree in Biomedical Engineering from Tulane University of Louisiana, having completed his dissertation research at the Biomedical Imaging Resource of the Mayo Clinic. He also obtained a fellowship from the General Electric Company to do research on digital image processing, during the fall of 1992, at the Magnetic Resonance Center of the GE Medical Systems (Waukesha, Wisconsin). He has been professor of the Universidad Nacional Autónoma de México, the Instituto Tecnológico de Colima and the University of Colima. Currently he is appointed as associate researcher at the Centro de Investigaciones en Optica A.C. in Leon Guanajuato, México.

    View full text