A Multifractal-Based Structural Representation of Retinal Images for Diabetic Retinopathy Analysis

Vision loss caused by diabetic retinopathy (DR) remains a serious public health challenge, driving the need for early and dependable automated detection tools. Conventional preprocessing techniques, such as CLAHE, histogram equalization, and gamma correction, mainly enhance image contrast without exploiting the structural complexity of retinal tissues. To overcome these limitations, the present work introduces a multifractal geometry-based framework for retinal image representation and automated DR analysis. The pipeline begins by removing blood vessels and the optic disc from fundus images, neutralizing the influence of normal anatomical structures and directing the analysis toward regions where pathological changes actually occur. A local multifractal analysis is then applied to extract two complementary descriptors: the generalized dimension, which tracks how structural complexity shifts across spatial scales, and the local Hölder exponent, which measures pointwise regularity and flags zones of irregular texture linked to tissue damage. Mapping both descriptors together produces a spatially resolved multifractal image that uncovers fine-grained microstructural differences separating damaged retinal tissue from healthy surrounding areas, something conventional enhancement methods consistently fail to reveal. This map is merged with the original fundus image to construct a two-channel input combining raw visual content with fractal-derived structural information. A convolutional neural network then processes both channels simultaneously to determine the DR severity grade of each image. Significant variations in multifractal maps across DR severity levels were confirmed by one-way ANOVA (p < 0.05), demonstrating that the method reliably captures disease-driven structural changes and provides discriminative features for grading. These outcomes highlight the capacity of multifractal modeling to support retinal image characterization and enable accurate automated DR diagnosis.