Improving microaneurysm detection in color fundus images by using context-aware approaches

https://doi.org/10.1016/j.compmedimag.2013.05.001Get rights and content

Abstract

In this paper, we present two approaches to improve microaneurysm detector ensembles. First, we provide an approach to select a set of preprocessing methods for a microaneurysm candidate extractor to enhance its detection performance in color fundus images. The performance of the candidate extractor with each preprocessing method is measured in six microaneurysm categories. The best performing preprocessing method for each category is selected and organized into an ensemble-based method. We tested our approach on the publicly available DiaretDB1 database, where the proposed approach led to an improvement regarding the individual approaches. Second, an adaptive weighting approach for microaneurysm detector ensembles is presented.The basis of the adaptive weighting approach is the spatial location and contrast of the detected microaneurysm. During training, the performance of ensemble members is measured with respect to these contextual information, which serves as a basis for the optimal weights assigned to the detectors. We have tested this approach on two publicly available datasets, where it showed its competitiveness compared without previously published ensemble-based approach for microaneurysm detection. Moreover, the proposed approach outperformed all the investigated individual detectors.

Introduction

Diabetic retinopathy (DR) is a serious eye disease that originates from diabetes mellitus and is the most common cause of blindness in the developed countries. Early treatment can prevent patients to become affected from this condition or at least the progression of DR can be slowed down. One of the earliest signs of DR are microaneurysms (MAs). Thus, it is essential to recognize this lesion in the fundus of the eye in time. Computer-aided MA detection is based on the detailed analysis of digital fundus images (see Fig. 1 for an example). MAs appear as small circular dark spots on the surface of the retina.

The detection of MAs highly depends on the characteristics of the imaging device and other image properties (e.g. type of compression). As a result, some MAs can be easily spotted on the background of the retina, while the recognition of others are more difficult. Besides image characteristics, the spatial location also has influence on the detection of MAs (e.g. proximity of vessel parts, etc.)

In [2], Niemeijer et al. distinguishes three categories based on visibility: subtle, regular and obvious. An example for this categorization can be seen in Fig. 2, Fig. 3. In the same study, they also investigate the detection of MAs near vessel. We extend these categorization with two additional categories with also taking into account the MAs which are detected on the macula and which are on the periphery of the image. Fig. 4, Fig. 5 show examples for the spatial categories. We also provide a computational approach to determine the characteristics of the MAs. In this paper, we propose two approaches exploiting this categorization to improve microaneurysm detection ensembles.

First, to recognize MAs in the different categories, we measure the effect of using different preprocessing methods. As we can see later on, a preprocessing method can enhance the detection rate in a few categories, but there is no single best performing preprocessing method for all. To overcome this difficulty, we propose a context-aware selection approach of preprocessing methods for MA candidate extraction.

Moreover, we also present an adaptive weighting approach for 〈preprocessing method, candidate extractor〉 (〈PP, CE〉 for short) ensembles [3]. In [4], we introduced 〈PP, CE〉 ensembles for MA detection, which are effective tools for increasing the sensitivity of microaneurysm detectors by fusing the detections of the candidate extractors applied after different preprocessing methods. In [5], we introduced a selection technique for 〈PP, CE〉 ensembles, which resulted in the first ranked microaneurysm detector in the Reintopathy Online Challenge [2]. In this paper, we present an adaptive weighting approach for 〈preprocessing method, candidate extractor〉 ensembles. This approach assigns an optimal weight to each member of the ensemble based on their performance of detecting MAs having different contrast and spatial locations. The experimental results show that this method is competitive with our former ensemble-selection approach [5].

The rest of the paper is organized as follows: in Section 2 we review the state-of-the-art for microaneurysm detectors. Section 3 introduces the context-dependent preprocessing method selection approach for MA candidate extractors. The concept of 〈PP, CE〉 ensembles is described in Section 4, as well as the proposed adaptive weighting approach. Section 5 is devoted to the methodology we used in this paper. In Section 6, we discuss our experimental results, while we draw conclusions in Section 7.

Section snippets

State-of-the-art

MA detection is based on the detailed analysis of digital fundus images. State-of-the-art detection approaches usually start with the preprocessing of images, which is followed by candidate extraction. Finally, the extracted candidates are classified as MAs or non-MAs. The reason to separate the latter two steps is that the pixel-wise classification of the whole image would be very resource-demanding.

The vast majority of microaneurysm detectors can be organized into two categories: the ones

Context-aware selection of 〈PP, CE〉 pairs

In this section, we describe our context-aware preprocessing method selection approach, which is based on learning. Thus, a training database with manually labelled MAs is needed. The creation of such database is explained in details in Section 3.1.

Adaptive weighting

In this section, we show a way to combine the output of the 〈PP, CE〉 pairs by weighting.

Methodology

In this section, we present the methodology used in this paper.

Context-aware selection

Table 2 contains the results of the training phase with the highest number of correctly recognized MAs are highlighted with the corresponding preprocessing method. As it can be seen, the CL, the IA and the VR preprocessing methods (see Section 5.1) are selected, while each performed better than the rest of the algorithms in two categories.

In Table 3, the results of the proposed approach and the individual methods can be seen. For fair comparison, we have selected that parameter setting for each

Conclusion

In this paper, we have presented two approaches to improve microaneurysm detection ensembles. First, we presented an approach to improve MA candidate extraction based on contextual information. That is, we categorized MAs based on their visibility and spatial location. We assumed that the use of different preprocessing methods helps candidate extractors to locate MAs belonging to different categories. To validate this assumption, we tested several preprocessing methods with a MA candidate

Acknowledgment

This work was supported in part by the project TAMOP-4.2.2.C-11/1/KONV-2012-0001 supported by the European Union, co-financed by the European Social Fund; the OTKA grant NK101680; TECH08-2 project DRSCREEN – Developing a computer based image processing system for diabetic retinopathy screening of the National Office for Research and Technology of Hungary (contract no.: OM-00194/2008, OM-00195/2008 and OM-00196/2008).

References (45)

  • B. Lay

    Analyse automatique des images angiofluorographiques au cours de la retinopathie diabetique

    (1983)
  • C.E. Baudoin et al.

    Automatic detection of microaneurysms in diabetic fluorescein angiographies

    Rev. Epidemiol Sante Publique

    (1984)
  • G.E. Oien et al.

    Diabetic retinopathy: automatic detection of early symptoms from retinal images

  • A. Mendonca et al.

    Automatic segmentation of microaneurysms in retinal angiograms of diabetic patients.

  • J.H. Hipwell et al.

    Automated detection of microaneurysms in digital red-free photographs: a diabetic retinopathy screening tool

    Diabetic Med

    (2000)
  • G. Yang et al.

    Algorithm for detecting micro-aneurysms in low resolution color retinal images

  • M.J. Cree et al.

    A fully automated comparative microaneurysm digital detection system

    Eye

    (1997)
  • L. Streeter et al.

    Microaneurysm detection in colour fundus images

  • A.D. Fleming et al.

    Automated microaneurysm detection using local contrast normalization and local vessel detection

    IEEE Trans Med Imaging

    (2006)
  • M. Niemeijer et al.

    Automatic detection of red lesions in digital color fundus photographs

    IEEE Trans Med Imaging

    (2005)
  • S. Ravishankar et al.

    Automated feature extraction for early detection of diabetic retinopathy in fundus images

  • J. Nayak et al.

    Automated identification of diabetic retinopathy stages using digital fundus images

    J Med Syst

    (2008)
  • Cited by (30)

    • Characteristic patch-based deep and handcrafted feature learning for red lesion segmentation in fundus images

      2023, Biomedical Signal Processing and Control
      Citation Excerpt :

      Niemeijer et al. [12] pre-processed the fundus images by removing the vessels and then used the k-NN classifier for red lesion detection. Antal et al. [13] proposed an ensemble-based MA detection algorithm, which uses adaptive weights for MA ensembles. Derwin et al. [14] presented a Local Neighborhood Differential Coherence Pattern (LNDCP) technique to extract features corresponding to MA segmentation using neural networks.

    • Cloud-based computer-assisted diagnostic solutions for e-health

      2020, Intelligent Data Security Solutions for e-Health Applications
    • L-Seg: An end-to-end unified framework for multi-lesion segmentation of fundus images

      2019, Neurocomputing
      Citation Excerpt :

      The fourth challenge is to distinguish lesions from other similar structures or noises. ( 5) Great difference usually exists in retinal fundus images obtained by different imaging devices, which increases the difficulty in detection of some lesions such as microaneurysms [3]. The fifth challenge is to build a flexible and robust model for different devices.

    • Automated segmentation of exudates, haemorrhages, microaneurysms using single convolutional neural network

      2017, Information Sciences
      Citation Excerpt :

      Walter et al. [33] on the other hand, used morphological processes and kernel density estimation to automatically segment micro-aneurysms. Quellec et al. [23] and Bae et al. [2] adopted template matching to perform the automated segmentation; Antal et al. [1] enlisted an ensemble of micro-aneurysms detectors to achieve the task. Almost all the above studies evaluated their output on the basis of lesion or image.

    View all citing articles on Scopus
    View full text