Micro-Lesion Segmentation in Diabetic Retinopathy Fundus Images

Diabetic Retinopathy (DR) is a chronic microvascular complication of diabe- tes mellitus, it’s one of the leading causes of preventable vision impairment worldwide and can result in blindness if it is not properly treated. The con- tinuous increase in blood sugar levels causes retina edema, leakage of blood and fluids in the retina, and development of new blood vessels, a situation called neovascularization. This pathological process destroys retina vessels, making it difficult to deliver vital nutrients such as oxygen, lipids, and pro- teins and therefore affecting visual function negatively. Early identification and monitoring of DR are essential to reduce the risk of irreversible retinal damage through timely treatment and follow-up. Retinal fundus photography is widely used in screening programs due to its non- invasive nature and relatively low acquisition cost, motivating the develop- ment of reliable computer-aided analysis tools to support clinical workflows. Most automated systems for DR focus on image-level grading, where a mo- del predicts a disease severity class. While grading is clinically valuable, it does not explicitly localize pathological patterns and therefore offers limited interpretability in terms of where and how lesions are distributed. Lesion segmentation, instead, aims to provide pixel-level maps of pathological re- gions and can support tasks such as quantitative lesion burden estimation, progression monitoring, and improved model explainability. This thesis addresses the problem of segmenting micro-lesions in retinal fun- dus images of patients affected by diabetic retinopathy. Micro-lesions include small pathological findings such as microaneurysms and small hemorrhages, which may occupy only a few pixels and often present limited contrast against the surrounding tissue. Accurate segmentation of such small targets is chal- lenging yet important, as these patterns are commonly associated with early stages of DR and can provide clinically meaningful signals for screening and longitudinal assessment.