Article Preview
TopIntroduction
Diabetic retinopathy (DR) is a complication of diabetes mellitus that may occur if the latter is left untreated. During DR, retina undergoes a group of lesions that gradually deteriorate it, being this condition the most frequent cause of blindness worldwide in individuals aged 20–74 years (Mahendran & Dhanasekaran, 2015). Therefore, detection in early stages is particularly important, as it helps ophthalmologists to administer appropriate treatments that improve patients health condition and hence protecting them from an irreversible damage, such as vision loss. The most common method employed by physicians for detecting DR is through the analysis of fundus photography, in which the disease shows up as hemorrhages, exudations and microaneurysms (MAs), being the last one the most important clinical finding in the diagnosis of DR, as they are the first lesion that appears during the early stages (Selcuk & Alkan, 2019; Sarhan et al., 2019). MAs appear as small red dots, and the main complication in their adequate detection is their similarity to vessel fragments (Bhatia et al., 2016). An example of MA in fundus image is shown in Figure 1.
Figure 1. Fundus image with presence of a MA
In common practice, the judgment by physicians is done visually based only on their experience, meaning that this process is time consuming, subjective and, as a result, this gives rise to differences in judgment. Besides, concerning the number of MAs found in fundus during the assessment, this helps to determine the severity of DR, however, the number may turn to be irrelevant, as the sole presence of one single MA is indicative of DR. Due to this, it is essential for computer-aided diagnosis (CAD) systems to detect MAs regardless their position or the amount of these. Inspired in avoiding subjectivity, speeding up the assessment process and increasing the sensitivity of fundus evaluation, several proposals have been suggested for detection and segmentation of MAs in fundus images.
TopBackground
Kumar and Kumar (2018) developed a method based on diverse image transformations that include principal component analysis (PCA) and contrast enhancement as pre-processing stage, and subsequently different regions of fundus photography are discarded, namely, exudates, blood vessels, optic disc and fovea, to remain only MAs, which are used to extract features and pass them to a support vector machine (SVM) for DR detection. Similarly, Joshi and Karule (2020) performed MAs detection by implementing image reconstruction over a pre-processed grayscale image obtained from green channel of RGB fundus image. Then, blood vessels and bright artifacts are removed and after connected component analysis, candidate MAs are detected. Lastly, after feature extraction, MAs are detected. The limitation of these approaches lies in the similarity of MAs with blood vessels, and for this reason, it is crucial a satisfactory blood vessels network segmentation. Furthermore, methods primarily based on image processing techniques are highly sensitive to non-uniform illumination occurred during fundus image acquisition, where the geometry of the eye causes internal reflection that contributes to the presence of shading artifacts (Manohar & Singh, 2018). For this reason, other researchers make the most of deep learning to achieve this task.