Abstract
Retinal damage caused by diabetes develops silently, and patients generally become aware of a problem only once their vision has deteriorated. Early automated detection therefore represents a concrete medical priority. Yet, a central question remains rarely addressed: are the image representations used by these systems actually designed for this purpose? Methods such as Contrast Limited Adaptive Histogram Equalization (CLAHE), histogram equalization, and gamma correction are widely used, but their primary objective is to improve image appearance rather than characterize the structural changes associated with different stages of diabetic retinopathy (DR). To address this limitation, this work introduces a multifractal geometry-based framework for retinal image representation and DR analysis. The framework consists of three main parts. First, normal anatomical structures, including blood vessels and the optic disc, are removed. Second, local variations in retinal structural regularity are estimated across different spatial scales. Third, multifractal maps are generated to represent local structural alterations in the retina. The removal of dominant anatomical structures focuses the analysis on the remaining retinal regions, while the multiscale analysis characterizes spatial variations in retinal structure associated with disease progression. Experiments performed on the Messidor-2 dataset showed stronger stage-related variations in multifractal maps than in CLAHE-enhanced images. The eta-squared analysis showed a stronger association between the extracted features and the severity of DR, with values reaching up to 0.955 for multifractal maps. In addition, the overall separation between DR severity stages improved from for CLAHE-enhanced images to 0.293 for multifractal maps, indicating a better distinction between severity stages. The key novelty lies in a lesion-centered multifractal representation that accounts for retinal structural complexity, thereby connecting multifractal analysis with intelligent systems for automated diabetic retinopathy grading.