Automatic detection of microaneurysms in color fundus images
Introduction
Diabetic retinopathy (DR) is one of the most serious and most frequent eye diseases in the world; it is the most common cause of blindness in adults between 20 and 60 years of age. DR is a complication of diabetes mellitus and although the disease does not necessarily coincide with vision impairment, about 2% of the diabetic patients are blind and 10% suffer from vision loss after 15 years of diabetes (Klein et al., 1995, Massin et al., 2000) due to DR complications.
Moreover, due to the increasing number of diabetic patients, the number of people affected by DR is expected to increase; DR has become a serious problem for our society and for the public health system. According to Lee et al. (2001), blindness due to diabetic eye disease produces costs of about 500 million dollars a year in the United States.
The main cause of the disease is the elevation of glucose in the blood, resulting in an alteration of the vascular walls whose first manifestations are microaneurysms (MA), tiny dilations of the capillaries. They can be seen as the first unequivocal sign of DR. This first abnormality does not affect the vision itself, but progression of DR to later stages leads to complications such as new vessels and macular edema, possibly leading to vision impairment and blindness. Both of these two main complications can be prevented by a proper treatment if the disease is detected early enough. Hence, early diagnosis of DR is essential for the prevention of vision impairment and blindness threatening a large number of persons in our society.
One of the easiest ways to diagnose DR is the analysis of color fundus images: their acquisition is cheap, non-invasive and easy to perform, and the most important lesions of DR are visible in this type of images. Although fluorescein angiographies (FA) allow detection of MA with a greater sensitivity, they are invasive and costly and therefore not adapted for screening purposes.
As DR is a silent disease in its first stages, i.e. many patients are not aware of its presence, it often remains undiagnosed until serious vision impairment occurs (Klein et al., 1992). Annual examinations of all diabetic patients are therefore highly recommended (Klein et al., 1995, Lee et al., 2001). The problem is however that this would produce tremendous costs, because the number of examinations would be enormous. Over and above that, there are not enough specialists to perform so many examinations; especially in rural areas, a mass-screening of DR is not feasible at present. An automatic pre-diagnosis based on image processing may help to overcome this problem by identifying automatically all “abnormal” retinas, i.e. all color images showing typical lesions for DR (Massin, 2002).
A further field of actual problems in the diagnosis of DR is the quantitative evaluation of different examinations in order to assess the evolution of the disease. At present, this is mostly done qualitatively. MA counts are a good indicator for the progression of DR (Klein et al., 1995, Hellstedt and Immonen, 1996, Massin et al., 2000) and can therefore be used for quantitative diagnosis. The aspect of quantification is extremely important for the efficiency assessment of a treatment or of a new drug. However, a manual comparison between different images is a very time-consuming and error-prone task. Once again, computer assisted diagnosis may help to overcome this problem (Zana, 1999).
As MA are the first unequivocal sign of DR as well as an indicator for its progression (Massin et al., 2000), their automatic detection plays a key role for both, mass-screening and monitoring and is therefore in the core of any system for computer assisted diagnosis of DR. In this paper, we present in detail a method for the automatic detection of MA, which has been developed in a mass-screening framework, without being limited to it. We do not present, in this article, a complete screening or monitoring system. A screening system usually contains other lesion detection algorithms, as well as an additional classification layer. A monitoring system requires, in addition to the automatic lesion detection, a robust registration algorithm.
The paper is structured as follows: First, we present former approaches for MA detection in color fundus images and angiographies. Then, we describe our method for automatic detection of MA based on criteria closings and kernel density estimation for automatic classification. Finally, we discuss the results of the algorithm.
Section snippets
State of the art
As the detection of MA is crucial to computer assisted diagnosis of DR, there is a large number of publications addressing this problem. Most of these publications deal with the detection of MA in fluorescein angiographies. The problem is similar: in the green channel of color images, MA appear as dark patterns, small, isolated and of circular shape. In FA, they appear as bright patterns and better contrasted, but the shape characteristics remain the same.
The first algorithm for the detection
An automatic method for the detection of MA in color fundus images
MA appear as small reddish isolated patterns of circular shape in color fundus images (Massin et al., 2000). They are characterized by their diameter which is always smaller than 125 μm. As they are situated on capillaries, and as capillaries are not visible in color fundus images, they appear as isolated patterns, i.e. disconnected from the vascular tree. MA have typically low contrast and they may be hard to distinguish from noise or pigmentation variations.
The pipeline of the method presented
Training set and test set
One of the major problems in the detection of MA is to establish a “gold standard”, i.e. a set of annotated images for learning (training set), and another for testing purposes (test set). MA can easily be confounded with other patterns; the manual results are therefore rather subjective. In this study, we have used a training set of 21 images. In all images of the training set, there was at least 1 MA. Moreover, 11 images showed exudates, 12 hemorrhages. The test set consisted of 94 images out
Conclusion
In this paper, we have presented a method of automatic detection of MA, based on diameter closing and kernel density estimation for automatic classification. We have tested this algorithm on a set of 94 images and we have obtained a sensitivity of 88.47% with 2.13 FP per image.
Although further evaluation of this algorithm on larger populations is still required, these results are satisfying. Any serious form of DR is unlikely to be missed and this should make ophthalmologists confident in using
References (37)
- et al.
Attribute openings, thinnings and granulometries
Computer Vision and Image Understanding
(1996) - et al.
A comparison of computer based classification methods applied to the detection of microaneurysms in opthalmic fluorescein angiograms
Computers in Biomedical Research
(1998) - et al.
Luminosity and contrast normalization in retinal images
Medical Image Analysis
(2005) - et al.
An image processing strategy for the segmentation and quantification of microaneurysms in fluorescein angiograms of the ocular fundus
Computers in Biomedical Research
(1996) - Autio, I., Borrás, J.C., Immonen, I., Jalli, P., Ukkonen, E., 2005. A voting margin approach for the detection of...
- Comaniciu, D., Ramesh, V., Meer, P., 2001. The variable bandwidth mean shift and data-driven scale selection. In: 8th...
- et al.
Spectral reflectance of the ocular fundus
Applied Optics
(1989) - et al.
Pattern Recognition and Scene Analysis
(1973) - Grisan, E., Ruggeri, A., 2005. A hierarchical Bayesian classification for non-vascular lesions detection in fundus...
- Fricout, G., Jeulin, D., 2004. Texture classification from their morphological properties. In: Proceedings of the 9th...
The Elements of Statistical Learning
Disappearance and formation rates of microaneurysms in early diabetic retinopathy
British Journal of Ophthalmology
The epidemiology of ocular problems in diabetes mellitus
S.S.Feman: Ocular Problems in Diabetes Mellitus
Retinal microaneurysm counts and 10-year progression of diabetic retinopathy
Archives of Ophthalmology
Comparison of diagnosis of early retinal lesions of diabetic retinopathy between a computer system and human experts
Archives of Ophthalmology
Rétinopathie Diabétique
Cited by (323)
Automatic detection of microaneurysms in fundus images based on multiple preprocessing fusion to extract features
2023, Biomedical Signal Processing and ControlMicroaneurysms detection in retinal images using a multi-scale approach
2023, Biomedical Signal Processing and ControlRecent trends and advances in fundus image analysis: A review
2022, Computers in Biology and MedicineEnhancing CNNs through the use of hand-crafted features in automated fundus image classification
2022, Biomedical Signal Processing and ControlCitation Excerpt :In addition to the publications mentioned above, there are many other papers in which the relevant features are extracted using conventional digital image processing tools. Walter et al. [5] used morphological processes and kernel density estimation to segment MAs and evaluated the input images based on the number of detected lesions. Quellec et al. [6] adapted optimal wavelet transform and template matching to perform automatic segmentation of MAs in retinal images.
Investigation of diabetic retinopathy level based on convolution neural network using fundus images
2024, Machine Learning Techniques for VLSI Chip Design