Elsevier

Pattern Recognition

Volume 45, Issue 7, July 2012, Pages 2832-2842
Pattern Recognition

Automatic optic disk boundary extraction by the use of curvelet transform and deformable variational level set model

https://doi.org/10.1016/j.patcog.2012.01.002Get rights and content

Abstract

Efficient optic disk (OD) localization and segmentation are important tasks in automated retinal screening. In this paper, we take digital curvelet transform (DCUT) of the enhanced retinal image and modify its coefficients based on the sparsity of curvelet coefficients to get probable location of OD. If there are not yellowish objects in retinal images or their size are negligible, we can then directly detect OD location by performing Canny edge detector to reconstructed image with modified coefficients. Otherwise, if the size of these objects is eminent, we can see circular regions in edge map as candidate regions for OD. In this case, we use some morphological operations to fill these circular regions and erode them to get final locations for candidate regions and remove undesired pixels in edge map. Since usually OD is surrounded by vessels, we choose the candidate region that has maximum summation of pixels in strongest edge map, which obtained by performing an appropriate threshold on the curvelet-based enhanced image, as final location of OD. Finally, the boundary of the OD is extracted by using level set deformable model. This method has been tested on different retinal image datasets and quantitative results are presented.

Highlights

► Optic disk (OD) segmentation is an important task in automated retinal screening. ► Curvelet transform is employed to get probable location of OD. ► Based on the size of yellowish objects two curvelet-based methods are proposed. ► We use some morphological operations to remove undesired pixels in edge map. ► The boundary of the OD is extracted by using level set deformable model.

Introduction

Optic disk (OD) is characterized as bright yellowish disk, from which, blood vessels and optic nerves emerge. It is a brighter region than the rest of the ocular funds and its shape is approximately round. The OD location can be used as a reference to identify the fovea in retinal images, for blood vessel tracking and other tasks. Many schemes have been proposed to detect OD. Tamura et al. [1] and Liu et al. [2] find the largest clusters of pixels with the highest intensities, while meets their difficulty when large hard exudates coexist in retinal image. Sinthanayothin et al. [3] identify the area with the highest intensity variation of adjacent pixels as OD. Li and Chutatape [4] use principal component analysis (PCA) to locate OD even for retinal images having bright lesions. Barrett et al. [5] propose applying Hough transform in order to locate OD. This method finds the circular shape with fixed radius in a thresholded edge image of the fundus and estimates the OD-contour by employing the Hough transform on detected edges. Niemeijer et al. [6] use a model of the geometrical directional pattern of the retinal vascular system, and implicitly embed the information on the OD position as the point of convergence of all vessels. Lalonde et al. [7] propose an OD detection method based on pyramidal decomposition and Hausdorff-based template matching. They estimate the OD-contour using Hausdorff-based matching between detected edges and a circular template. In general, the individual attributes are not sufficient to decide final OD location. For instance, in uneven illumination conditions or in the presence of bright lesions brightness, it is necessary to combine all attributes to handle all possible variants of a retinal image.

In this paper, we present a new curvelet-based method for OD detection which is a hybrid approach that characterizes OD using intensity and vessel structures. The pipeline of the presented method is illustrated in Fig. 1 (the left column shows the proposed algorithm for extraction of blood vessels, and the right column shows the pipeline of proposed algorithm for determination of OD candidate locations). Candidate regions for OD location are determined by modification of curvelet coefficients of enhanced image, and performing Canny edge detector and some morphological operations on reconstructed image with this modified coefficients. We choose the candidate region that has maximum summation of pixels in strongest edge map, as final location of OD. After detection of OD location, we use variational level set deformable model and apply the initial level set contour around the detected OD center and extract the whole OD boundary.

The paper is organized as follows. After a short review on digital curvelet transform (DCUT) via wrapping, the proposed method in this paper is introduced in Section 2. Our curvelet-based technique includes 4 steps including: (1) contrast improvement and non-uniform illumination correction, (2) curvelet-based determination of candidate regions, (3) OD detection, and (4) OD boundary extraction. In Section 3, the results and discussion are presented, and finally this paper is concluded in Section 4.

Section snippets

Proposed method

DCUT was introduced to represent edges and other singularities along curves much more efficiently than other traditional transforms such as Fourier transform, ridgelet transform and wavelet transform [8], [9]. Ridgelet analysis may be constructed as wavelet analysis in the Radon domain. The Radon transform translates singularities along straight lines into point singularities. Then, the wavelet transform can be used to effectively handle these point singularities. In images, edges are typically

Results

The proposed method for detection of OD location has been evaluated on three publicly available datasets, the STARE [13], DRIVE [24], and DIARETDB1 [25]. We considered the OD detected center within 7 pixels (with average OD within 40×40 pixels) of its actual position as successful detection.

The quantitative results obtained from the proposed method are shown in Table 1.

The experimental result shows that the proposed algorithm is robust to variable appearance of retinal images and can give the

Conclusion

Because the brightness, contrast and color of OD vary a lot among different patients, some previous methods for OD detection would not work well in all the images used in clinical environment. In this paper, we introduced a new method for OD detection based on the DCUT. In the curvelet domain, some features like exudates and OD in retinal images are correspond to large coefficients in curvelet domain. So, we perform Canny edge detector to extracted image from strengthen curvelet coefficients

Acknowledgement

The authors would like to thank Dr Alireza Dehghani from Ophthalmology Department, Isfahan University of Medical Sciences for his contribution in data collection and evaluation of results.

Hossein Rabbani was born in Iran in 1978. He received the B.Sc. degree in Electrical Engineering (Communications) from Isfahan University of Technology, Isfahan, Iran, in 2000 with the highest honors, and the M.Sc. and Ph.D. degrees in Bioelectrical Engineering in 2002 and 2008, respectively, from Amirkabir University of Technology (Tehran Polytechnic), Tehran, Iran. From January 2007 to July 2007, he was with the Department of Electrical and Computer Engineering, Queen's University, Kingston,

References (27)

  • S. Tamura et al.

    Zero crossing interval correction in tracking eye-fundus blood vessels

    Pattern Recognition

    (1988)
  • S. Osher et al.

    Level set methods: an overview and some recent results

    Journal of Computer Physics

    (2001)
  • Z. Liu, O. Chutatape, S.M. Krishnan, Automatic image analysis of fundus photograph, in: Proceedings of the 19th Annual...
  • C. Sinthanayothin et al.

    Automated location of the optic disk, fovea, & retinal blood vessels from digital color fundus images

    British Journal of Ophthalmology

    (1999)
  • H. Li et al.

    Automated feature fxtraction in color retinal images by a model based approach

    IEEE Transactions on Biomedical Engineering

    (2004)
  • S.F. Barrett et al.

    Employing the hough transform to locate the optic disk

    Biomedical Sciences Instrumentation

    (2001)
  • M. Niemeijer et al.

    Segmentation of the optic disc, macula and vascular arch in fundus photographs

    IEEE Transactions on Medical Imaging

    (2007)
  • M. Lalonde et al.

    Fast and robust optic disk detection using pyramidal decomposition and Hausdorff-based template matching

    IEEE Transactions on Medical Imaging

    (2001)
  • E. Cand'es et al.

    Fast discrete curvelet transforms

    Multiscale Modeling and Simulation

    (March 2006)
  • D. Donoho, M.R. Duncan, Digital Curvelet Transform: Strategy, Implementation, Experiments, Technical Report, Stanford...
  • J.-L. Starck et al.

    Gray and color image contrast enhancement by the curvelet transform

    IEEE Transactions on Image Processing

    (2003)
  • J. Ma et al.

    The curvelet transform

    IEEE Signal Processing Magazine

    (2010)
  • E. Pisano et al.

    Contrast limited adaptive histogram equalization image processing to improve the detection of simulated spiculations in dense mammograms

    Journal of Digital Imaging

    (1998)
  • Cited by (0)

    Hossein Rabbani was born in Iran in 1978. He received the B.Sc. degree in Electrical Engineering (Communications) from Isfahan University of Technology, Isfahan, Iran, in 2000 with the highest honors, and the M.Sc. and Ph.D. degrees in Bioelectrical Engineering in 2002 and 2008, respectively, from Amirkabir University of Technology (Tehran Polytechnic), Tehran, Iran. From January 2007 to July 2007, he was with the Department of Electrical and Computer Engineering, Queen's University, Kingston, ON, Canada, as a Visiting Researcher. Then he joined Department of Biomedical Engineering and Medical Image & Signal Processing Research Center, Isfahan University of Medical Sciences as an Assistant Professor. He is also currently a Postdoctoral Research Scholar at University of Iowa.

    His main research interests are medical image analysis, multidimensional signal processing, multi-resolution transforms, probability models of sparse domain's coefficients and image restoration.

    View full text