From a medical point of view, there are three main domains of algorithms which can be conceived for the improvement of diagnosis of diabetic retinopathy:
Figure 7.2: Image of the fundus.
1. Image enhancement: Images taken at standard examinations are often noisy and poorly contrasted. Over and above that illumination is normally not uniform. Techniques improving contrast and sharpness and reducing noise are therefore required.
2. Mass screening: Computer-assisted mass screening is certainly the most important task to which image processing can contribute. We have already seen that the blinding complication of diabetic retinopathy can be inhibited by early treatment. However, as vision normally alters only in the later stages of the disease, many patients remain undiagnosed in the earlier stages of the disease. Hence, mass screening of all diabetic patients would help to diagnose this disease early enough. Unfortunately, this approach is not very realistic, taking into consideration the large number of diabetic patients compared to a lack to specialists. Computer assistance could make mass screening a lot more efficient.
3. Monitoring: Comparing images taken at different examinations allows one to evaluate a treatment or new therapeutics. However, it is a time-consuming task and open to human error. Computer-based comparison including automatic registration and evaluation of changes between images could deliver a precious tool for monitoring the disease.
We have seen in this introductory section that diabetic retinopathy is a real problem for a high number of diabetic patients and even for our health systems. We have also seen possible approaches of computer assistance that may help to overcome actual problems in its diagnosis.
Of course, giving detailed solutions to all these problems would go over the scope of this paper. After having analyzed the nature of color in fundus photographs and after having given a short introduction into mathematical morphology, a nonlinear image-processing technique our algorithms are mainly based on, we will describe in detail some algorithms within this framework: We will present an algorithm for image enhancement, algorithms for the detection of the vessels and the optic disk, and finally algorithms for the detection of characteristic lesions like microaneurysms and exudates. These segmentation algorithms are essential for computer-assisted screening and monitoring systems.
Was this article helpful?