Retinal vessel segmentation plays a crucial role in diagnosing several serious conditions, including diabetic retinopathy, glaucoma, and cardiovascular disease. In this work, we present a hybrid framework that brings together fractal geometry, nonlinear diffusion, and deep learning techniques. Our approach constructs a 12-channel feature representation that integrates the original RGB channels with multi-scale vesselness filters, wavelet energy, enhancement through partial differential equations (PDEs), and fractal descriptors—specifically fractal dimension and lacunarity. We use a U-Net architecture enhanced with CBAM attention mechanisms to process these features, training it with a composite loss function that incorporates Dice loss, centerline-Dice, and a novel fractal consistency term. Testing on the DRIVE dataset, we achieved an 86.91% Dice coefficient, representing a statistically significant improvement over existing baselines. The framework is computationally efficient with only 5.65 M parameters and 75 ms inference time, which makes it practical for clinical deployment. Addi-tionally, the extracted physics-informed features are interpretable and have been validated through quantitative attention analysis and feature importance studies.