Biomedical image quality improvements with Attention Mechanisms and Deep Residual Learning

Abstract

Image segmentation plays a key role in biomedical imaging, allowing different structures or regions of interest, such as organs, tissues, or blood vessels, to be clearly identified for further analysis and diagnosis. Segmentation of retinal vessels in fundus images is particularly challenging due to low contrast and complex vascular structure, which complicates the diagnosis of ophthalmic diseases. In this study, we present a state-of-the-art deep learning approach to improve biomedical image quality and segmentation accuracy using Attention U-Net++ combined with the ResNet support network. Attention mechanisms enhance the model"s ability to focus on fine details of blood vessels, while deep residual learning ensures the stability of deep architecture learning. The method is evaluated on a fundus image dataset, achieving a Dice score of 0.8694 and an Intersection over (IoU) score of 0.7690, demonstrating competitiveness with state-of-the-art methods. The results of the visual analysis emphasize the model\"s ability to accurately delineate vessels and indicate areas for improvement, namely the development of methods for finding small vessels. The results demonstrate the potential of attention and residual learning mechanisms to improve biomedical image analysis, offering a reliable tool for clinical applications in the diagnosis of retinal diseases.