MRUF-Net: A U-Net Based Retinal Vessel Segmentation with Retinex Contrast Enhancement and Focal-Weighted loss Function

The retinal vessel network can provide valuable information about both ocular and systemic health problems. Therefore, accurate segmentation of the retinal vessels has been a central focus of researchers. However, due to the complex, varied, and heterogeneous structure of vessels, it remains a significant challenge. Despite advancements in deep learning approaches, many state-of-the-art approaches in this domain rely heavily on architectural innovations, while neglecting the important role of image processing techniques such as contrast enhancement and model training guidelines such as loss function design. This paper proposes a comprehensive solution, called MRUF-Net, integrating multiscale Retinex with colour preservation (MSRCP) for contrast enhancement, U-Net deep model with both standard and lightweight (with half-depth) configurations, and a hybrid loss function that combines focal loss with weighted binary cross-entropy (WBCE) for retinal vessel segmentation. Experimental segmentation results on three benchmark retinal datasets of DRIVE, CHASE_DB1, and STARE demonstrate that the proposed approach consistently outperforms existing techniques, even with reduced model complexity, underscoring the effectiveness of targeted image preprocessing and optimized loss formulations in enhancing retinal vessel segmentation performance.