Automated detection of microaneurysms in retinal images using convolutional neural networks

Abstract

Microaneurysms (MAs) are one of the earliest and most critical newlineindicators of Diabetic Retinopathy (DR), a serious consequence of diabetes that newlinemay result in vision loss. MAs are small, localized dilations of retinal capillaries newlinecaused by prolonged hyperglycemia, which weakens the capillary walls. These newlineMAs appear as tiny red dots on the retinal surface and can rupture, leading to newlineintra-retinal hemorrhages and further retinal damage. However, accurately newlineidentifying MAs amidst the complex background of retinal images poses newlinesignificant challenges. The small size of MAs, combined with their subtle newlineappearance, makes them difficult to distinguish from other retinal features. newlineTraditional methods for detecting microaneurysms (MAs) in Diabetic Retinopathy newline(DR) including morphological operations, intensity-based filtering, and classical newlinemachine learning techniques face notable challenges that hinder accurate newlinediagnosis. These techniques often struggle with the small size (10 100 and#956;m) and newlinelow contrast of MAs, which makes them difficult to distinguish from the newlinesurrounding retinal tissue, especially in low-quality or low-resolution fundus newlineimages. Moreover, complex backgrounds with overlapping blood vessels and newlineanatomical features can obscure MAs, leading to false positives or missed newlinedetections. The variability in MA shape, size, and intensity across patients, along newlinewith the presence of image noise and artifacts, further reduces the robustness of newlinetraditional methods. Classical classifiers such as SVMs, Random Forests and newlinePrincipal Component Analysis (PCA) often fail to capture the subtle characteristics newlineand spatial context of MAs, resulting in poor generalization and increased newlinecomputational overhead. To address these challenges, this thesis introduces two newlineadvanced methodologies designed to enhance both the precision and effectiveness newlineof MAs detection in DR. newline