Microaneurysm detection in color eye fundus images for diabetic retinopathy screening
Introduction
Diabetes mellitus (DM), a long-term condition characterized by high levels of glucose in the blood, is one of the fastest growing global health issues of the 21st century. According to the International Diabetes Federation estimates, diabetes affects approximately 463 million people aged 20–79 years worldwide in 2019 and, if the current trend continue, that number will increase to 700 million by 2045 [1].
As one of the most common complications of DM, DR has also become more prevalent over the years [2]. Currently, DR affects more than one third of the diabetic population and it is one of the leading causes of vision loss in working-age adults [3]. However, in most cases it may be avoidable through an early diagnosis and an appropriate treatment. Since the first stages of DR are often asymptomatic, regular eye examination should be performed in diabetic people in order to make possible its detection at an early stage.
A correct examination of the retina implies the acquisition of retinal photographs using a fundus camera. Those images are analysed afterwards by an ophthalmologist, who looks for retinal lesions in order to conclude if DR is present or not. According to the findings, the ophthalmologist infers about the degree of severity and decides whether the patient should be referred for treatment or not [4]. Microaneurysms, hemorrhages, exudates and neovessels are examples of retinal lesions in DR patients.
Besides eye fundus image analysis being a very time-consuming and resource-demanding task, most images acquired in a DR screening program do not present signs of the disease (Fig. 1a). In order to alleviate the burden of the ophthalmologists and increase the efficiency of DR screening, several research teams have focused on the development of fully automated methods for detection of retinal lesions in fundus images and DR diagnosis over the last years [5,6].
Most of the methods for DR lesion detection are focused on the detection of MAs, because they are one of the first clinical signs of the disease and, therefore, crucial for its early detection, namely in screening programs. Although MAs appear in the retinal photographs as “isolated, spherical, red dots” [4] (Fig. 1b), MA detection is a challenging task due to the expected heterogeneity of the images, poor quality of many images and similarity between MAs and other retinal structures or lesions (e.g. small hemorrhages) [7].
The conventional approaches developed for MA detection in eye fundus images usually comprise three main steps: preprocessing, candidate extraction and candidate classification [8]. Although different preprocessing techniques have been applied for contrast enhancement and shade correction [9], most of these methods use templates/filters which are static in terms of size and/or shape for MA enhancement and candidate extraction [10,11], limiting the type of MAs detected.
Contrary to these filters, sliding band filters (SBFs) rely on a support region with a band whose position may vary in each direction [12] and therefore allow to detect convex objects in a wider range of sizes and shapes. Although the responses of this kind of filters have been used as features for MA candidate classification [13], the use of SBFs for MA enhancement was not explored yet.
In this work, we propose a new methodology for MA detection in fundus images, using a SBF for MA enhancement before the extraction of potential candidates. Since these filters are based on local gradient convergence, they can handle problems related to illumination variations, noise and low contrast [14]. A different combination of features, including the filter responses, intensity, contrast and shape-based features, is also used for a better discrimination between the true microaneurysms and false detections. To validate the methodology, our approach was tested in the most cited public datasets with MA annotations, as well as in a private dataset containing images acquired in the context of a DR screening program. The feasibility of integrating the developed methods in a computer aided diagnosis (CAD) system for DR detection is also evaluated in a large set of eye fundus images.
The rest of the paper is organized as follows: Section 2 gives a short review of the existing methods for MA detection in retinal images, Section 3 presents some details about the SBF as well as the advantages to applying it for MA detection and Section 4 presents the methodology herein proposed for candidate extraction and classification. Then, Section 5 provides a description of the datasets used for evaluating the developed methods and Section 6 presents some details about parameter settings. In Section 7, the obtained results are shown and compared with the performance of some state-of-the-art methods and the agreement between different ophthalmologists. Finally, the main conclusions of this work are stated in Section 8.
Section snippets
Review of the existing methods for microaneurysm detection
During the last two decades, many researchers have been focused on the development of CAD systems for DR. Most of the studies are related to the detection of retinal landmarks (such as blood vessels, optic disc and fovea) or lesions (MAs, hemorrhages, exudates and cotton wool spots) [6,15]. Here, only the methods proposed for MA detection will be reviewed.
Although the first studies in the field had been developed using fluorescein angiograms (that present a good contrast between MAs and
Sliding band filter
The SBF was primarily introduced by Pereira et al. [12] for the enhancement of objects with a convex shape. As an LCF, the SBF is based on the maximization of the convergence index at each image point and can be defined bywhere N represents the number of support region lines, and the inner and outer sliding limits of the band, respectively, the angle of the gradient vector at the point m pixels away from in
Methodology
As most of the methods described in the literature, the methodology herein proposed for MA detection can be divided into three sequential steps: image preprocessing, candidate extraction and candidate classification (Fig. 3). In the image preprocessing step, a de-noising technique is applied in order to improve the quality of the input image and MAs are enhanced using the SBF. Afterwards, several operations are applied to the preprocessed image in order to find the initial set of MA candidates.
Datasets
In order to validate the methodology herein proposed for MA detection, four different datasets were used. Three of them are publicly available and are commonly used for the development and evaluation of MA and DR detection methods. The private dataset was obtained in a real scenario of DR screening and, therefore, comprises images with different resolutions and quality. The main specifications of the datasets are summarized in Table 2.
Parameter tuning
As mentioned in subsection 4.1, the proposed method applies a Gaussian filter and a SBF to the input image, in the preprocessing stage, in order to remove noise and enhance the MAs before candidate extraction. The filters’ parameters must therefore be chosen according to the size of the MAs to be detected (in pixels), which is mainly dependent on the image resolution and the FOV angle.
In this study, the filters’ parameters were empirically tuned for a subset of images of the ROC training set
Results and discussion
In this section, the algorithm and metrics used for evaluating the performance of the method, at the lesion and image-level, are firstly presented. Then, the results obtained in the public datasets are compared with those achieved by some state-of-the-art methods. For the SCREEN-DR database, we compare the performance of the developed method with each one of the medical experts who annotated the images.
Conclusions
In the research herein presented, a fully automated method for MA detection in eye fundus images is proposed. It comprises three main steps: preprocessing, MA candidate extraction and classification. Initially, a sliding band filter is applied to enhance the MAs and the potential candidates are detected by combining the filter responses with intensity and shape information. Then, 31 features are extracted from the MA candidates and a RUSBoost classifier is used for discriminating MA from non-MA
Declaration of interests
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work is financed by the ERDF – European Regional Development Fund through the Operational Programme for Competitiveness and Internationalisation – COMPETE 2020 Programme, and by National Funds through the FCT – Fundação para a Ciência e a Tecnologia within project CMUP-ERI/TIC/0028/2014.
Tânia Melo is funded by the FCT grant SFRH/BD/145329/2019.
We would also like to thank the ophthalmologists from Centro Hospitalar Universitário de São João who annotated the images of the SCREEN-DR database.
References (58)
- et al.
A review on automatic analysis techniques for color fundus photographs
Comput. Struct. Biotechnol. J.
(2016) - et al.
Diabetic retinopathy techniques in retinal images: a review
Artif. Intell. Med.
(2019) - et al.
IDRiD: diabetic retinopathy – segmentation and grading challenge
Med. Image Anal.
(2020) - et al.
Automatic detection of microaneurysms in retinal fundus images
Comput. Med. Imag. Graph.
(2017) - et al.
Detection of microaneurysms using multi-scale correlation coefficients
Pattern Recogn.
(2010) - et al.
Computer-aided diagnosis of diabetic retinopathy: a review
Comput. Biol. Med.
(2013) - et al.
Algorithms for red lesion detection in diabetic retinopathy: a review
Biomed. Pharmacother.
(2018) - et al.
An image-processing strategy for the segmentation and quantification of microaneurysms in fluorescein angiograms of the ocular fundus
Comput. Biomed. Res.
(1996) - et al.
Automatic detection of microaneurysms in color fundus images
Med. Image Anal.
(2007) - et al.
Automated detection of microaneurysms using scale-adapted blob analysis and semi-supervised learning
Comput. Methods Progr. Biomed.
(2014)