Application of Mathematical Morphology
to the diagnosis of Diabetic Retinopathy using color fundus images
This thesis deals with the automatic analysis of color fundus images and with its application to the diagnosis of Diabetic Retinopathy.
Diabetic Retinopathy - a complication of diabetes mellitus - is a severe and wide-spread eye disease: it is the leading cause of legal blindness for the working age population in western countries. For the diagnosis of Diabetic Retinopathy, digital color fundus images are becoming increasingly important. This fact opens the possibility of applying image processing techniques in order to facilitate and improve diagnosis in different ways. First, manual analysis can be improved by using image enhancement methods. Second, the main problem of diagnosing Diabetic Retinopathy too late for an optimal treatment may be resolved by a computer assisted mass-screening approach. Third, image registration techniques can be used in order to compare images taken at different examinations quantitatively. This comparison is essential for a monitoring of the disease. These three ways of diagnosis improvement have been studied in this thesis, and algorithms within this framework have been developed.
We present new image enhancement
methods like contrast enhancement and shade correction.
Automatic methods for mass-screening and monitoring of Diabetic Retinopathy are based on feature detection, i.e. the detection of the main anatomical structures of the retina, and the detection of lesions. The detection of anatomical features is essential for any system of analysis of fundus images as they allow one to classify lesions and to find landmarks for image registration. We present and discuss new approaches for the detection of the vascular tree, the optic disc and the macula.
Lesion detection is also essential for mass-screening and monitoring purposes. We present a new method for the detection of microaneurysms. Microaneurysms are the first unequivocal sign of Diabetic Retinopathy and their number is a good indication for the progression of the disease. Hence, their diagnostic value is particularly important, and their detection essential for both the mass-screening and the monitoring task. As Diabetic Retinopathy is a silent disease, it is important to detect the transition to the stage of complication. The two complications that are the reason for vision alteration and blindness are neovascular responses to local ischemia and macular edema. It is difficult to detect these complications directly using color fundus images. However, a high number of hemorrhages are an important indication for local ischemia which is the reason for new vessels; hard exudates are a major hallmark for retinal edema. We present methods for the automatic detection of these two lesions.
With the algorithms developed in
this thesis, it is possible to conceive diagnostic tools that may play a
major role in mass-screening and monitoring of Diabetic Retinopathy. In combination
with a system based on telemedecine, these tools may allow an efficient mass-screening,
even in countries that show a considerable lack of specialists in ophthalmology.
A further domain of application may be medical pharmaceutical research, as
they allow a quantitative and reproducible assessment of the evolution of
the disease after treatment.