Elsevier

Medical Image Analysis

Volume 13, Issue 4, August 2009, Pages 650-658
Medical Image Analysis

Retinal image analysis based on mixture models to detect hard exudates

https://doi.org/10.1016/j.media.2009.05.005Get rights and content

Abstract

Diabetic Retinopathy is one of the leading causes of blindness in developed countries. Hard exudates have been found to be one of the most prevalent earliest clinical signs of retinopathy. Thus, automatic detection of hard exudates from retinal images is clinically significant. In this study, an automatic method to detect hard exudates is proposed. The algorithm is based on mixture models to dynamically threshold the images in order to separate exudates from background. A postprocessing technique, based on edge detection, is applied to distinguish hard exudates from cotton wool spots and other artefacts. We prospectively assessed the algorithm performance using a database of 80 retinal images with variable colour, brightness, and quality. The algorithm obtained a sensitivity of 90.2% and a positive predictive value of 96.8% using a lesion-based criterion. The image-based classification accuracy is also evaluated obtaining a sensitivity of 100% and a specificity of 90%.

Introduction

Retinal images are widely used by ophthalmologists and primary care physicians for the screening of epidemic eye diseases, such as Diabetic Retinopathy (DR). DR is one of the leading causes of blindness and vision defects in developed countries. Fundus images permit a high quality permanent record of retinal fundus for detecting early signs of DR and monitoring its progression. Moreover, their digital nature allows automatic analysis to reduce the workloads of the ophthalmologists and the health costs in the screening of the disease.

Early detection of DR is crucial for the prevention of visual loss. Hard exudates (HEs) are one of the most prevalent signs in the earliest stages of the disease (Singer et al., 1992). Additionally, they represent the most specific maker for the presence of co-existent retinal oedema, the major cause of visual loss in the non-proliferative forms of DR (Singer et al., 1992). Automatic exudates detection is a difficult task due to the uneven illumination, poor contrast and colour variation of the retinal images. Several attempts have been made to segment this sign from the retinal background.

Grey level thresholding is the simplest exudate segmentation method (Kavitha and Devi, 2005, Liu et al., 1997, Philips et al., 1993, Ward et al., 1989). However, the automatic selection of a threshold is difficult due to the uneven illumination of the HEs. A global threshold is set manually for each image in (Ward et al., 1989), whereas a local threshold is used in (Philips et al., 1993) for different regions, which are selected by the user. A local dynamic thresholding algorithm is proposed in (Liu et al., 1997), based on the histogram shape but no results are reported. In another approach, a multilevel thresholding of the histogram is applied to segment exudates (Kavitha and Devi, 2005). Although a sensitivity of 100% and a mean number of 0.1 false positives per image were obtained, the resulting algorithm was tested using only ten images.

The grey level variation of the exudates using mathematical morphology is also exploited for HE detection (Walter et al., 2002), obtaining a sensitivity of 92.8% and a positive predictive value of 92.4% in a small dataset of 30 retinal images.

Region growing algorithms segment retinal images based on the homogeneity of the exudates illumination (Sinthanayothin et al., 2002) or combined with edge detection (Li and Chutatape, 2004). These techniques have the drawback of being computationally intensive.

Clustering algorithms (Hsu et al., 2001), statistical classification (Goh et al., 2000, Wang et al., 2000), supervised approaches (Niemeijer et al., 2007) and neural network (Gardner et al., 1996, Osareh, 2004, Zhang and Chutatape, 2005) have been also attempted to detect HE. They have used different exudate features (illumination, contrast, colour, etc.) guaranteeing the detection of exudates at the expense of simplicity.

Our purpose is to select a threshold that allows segmentation of the exudates from the background. A fixed value based on the exudates grey level cannot be assigned for this threshold because of the wide variability in illumination and contrast from image to image, strongly correlated to subject’s intrinsic characteristics. We propose a innovative segmentation approach based on a statistical mixture model based clustering, which allows a robust segmentation of the image foreground in a totally unsupervised manner. In contrast to supervised algorithms (e.g. Niemeijer et al., 2007), there is no need of training phase or feature extraction procedure, reducing the computation costs. In other approaches (e.g. Osareh, 2004) the segmentation relies on the study of the histogram shape, which is highly influenced by outliers. The proposed method, by contrast, can deal with outlying observations, obtaining a robust separation of the foreground and background scenes and, specifically, a segmentation of hard exudates. A postprocessing technique, based on edge detection, is then applied to distinguish HEs from other bright lesions, such as cotton wool spots, and from other bright elements that are detected after the segmentation process, such as artefacts along large blood vessels due to light reflection. The resulting algorithm achieves a satisfactory performance with independence of the variable appearance of retinal fundus images, an important factor in clinical environment.

Section snippets

Mixture models

Mixture models (MMs) are a powerful semi-parametric statistical technique for estimating probability densities, such as histograms (McLachlan and Peel, 2000, Titterington et al., 1985). Due to their advantages over nonparametric and parametric methods, MMs are being increasingly exploited in medical image analysis to model pixel values as a combination of different populations mixed in varying proportions (Frosio et al., 2006, Noe and Gee, 2001, Osareh, 2004).

Let Yj be a p-dimensional random

Methods

Exudates segmentation is obtained by five sequential steps. As stated in (Walter et al., 2002), the exudates appear more contrasted in the green component IG of the RGB colour model. First, IG is enhanced to obtain a histogram where HEs can be better segmented from the background. The histogram of the enhanced image is modelled using a MM and a dynamic threshold is set based on the statistical information obtained. Then, the optic disk is automatically localized and masked out. Finally, a

Retinal images and system performance evaluation

We assessed the performance of our algorithm on retinal images provided by the Instituto de Oftalmobiología Aplicada at University of Valladolid, Spain. The image data set consists of 106 images taken with a TopCon TRCNW6S non-mydriatic retinal camera at 45° field of view. Image resolution was 576 × 768 pixels in 24 bit JPEG format. The images came from a clinical set of diabetic patients who were referred to the ophthalmologist for further examination. Retinopathy grading was performed on these

Results

For each retinal images of the database, we modelled the histogram with a 3-component MM obtaining an average absolute error of 5.1 ± 2.9%.

We tested the complete algorithm using a lesion-based criterion over the test set. Fig. 5a shows the sensitivity–predictivity curve varying Thpp. The best overall performance was obtained at Thpp = 0.12 with a mean sensitivity of 90.2% and a mean predictive value of 96.8%. The effect of varying Thpp on the algorithm performance can be analyzed from this graph:

Design considerations

The proposed exudate detection algorithm using MMs is an innovative thresholding method which sets the threshold based on the robust estimation of the histogram. Modelling each histogram with a different MM, we obtain a dynamic threshold for each image.

Preprocessing stage is crucial for the algorithm success due to the variability in colour, illumination and contrast of the retinal images. The proposed preprocessing approach obtains grey level normalization at the same time as contrast

References (27)

  • M. Foracchia et al.

    Luminosity and contrast normalization in retinal images

    Medical Image Analysis

    (2005)
  • N.P. Ward et al.

    Image analysis of fundus photographs – the detection and measurement of exudates associated with diabetic retinopathy

    Ophthalmology

    (1989)
  • British Academic Association

    Retinal Photography Screening for Diabetic Eye Disease

    (1997)
  • S. Chaudhuri et al.

    Detection of blood vessels in retinal images two dimensional matched filters

    IEEE Transactions on Medical Imaging

    (1989)
  • A.P. Dempster et al.

    Maximum likelihood from incomplete data via the EM algorithm

    Journal of the Royal Statistical Society

    (1977)
  • I. Frosio et al.

    Enhancing digital cephalic radiography with mixture models and local gamma correction

    IEEE Transaction on Medical Imaging

    (2006)
  • K. Fukunaga

    Statistical Pattern Recognition

    (1990)
  • G. Gardner et al.

    Automatic detection of diabetic retinopathy using an artificial neural network: a screening tool

    British Journal of Ophthalmology

    (1996)
  • K.G. Goh et al.

    ADRIS: an automatic diabetic retinal image screening system

  • Hsu, W., Pallawala, P.M.D.S., Lee, M.L., Eong, K.-G., 2001. The role of domain knowledge in the detection of retinal...
  • A. Jain

    Fundamentals of Digital Image Processing

    (1989)
  • Kavitha, D., Devi, S.S., 2005. Automatic detection of optic disk and exudates in retinal images. In: Proceedings of the...
  • H. Li et al.

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

    IEEE Transactions on Biomedical Engineering

    (2004)
  • Cited by (0)

    View full text