<p>In order to accurately diagnose ocular disorders DR, Glaucoma and AMD with computer assistance, it is essential to segment retinal structures in fundus images. Although deep learning-based segmentation models have made significant strides, current approaches have significant drawbacks: The inability to adaptively focus on lesion-relevant regions, the inability to capture fine vascular features, and the reliance on manually created or annotated lesion priors are common problems with traditional U-Net variations. Additionally, the channel and spatial attention mechanisms used in previous studies often struggle with effective feature recalibration and are unable to simulate long-range dependencies, which results in segmentation accuracy which is not ideal. We propose QBFAM-U-Net, a novel Quad Branch Fusion Attention Module (QBFAM) combined with U-Net, to overcome these drawbacks. The QBFAM presents a residual learning framework with a cascaded channel–spatial attention mechanism. In particular, our design makes use of (i) skip pathways to reduce vanishing gradients and stabilise deep network training, (ii) multi-branch dilated convolutions for efficient receptive field enlargement without resolution loss, and (iii) adaptive attention recalibration to specifically highlight vessel-like and lesion-aware features. QBFAM greatly enhances feature representation across scales and preserves spatial details when included into the U-Net encoder-decoder framework. Comprehensive tests on benchmark datasets including CHASE_DB1, DRIVE, and STARE are utilized for the experiment. The evaluation criteria, such as Accuracy, Precision, Specificity, Sensitivity, F1 Score, and AUC are used. This development demonstrates QBFAM-U-Net's potential as a reliable and effective retinal image analysis tool, offering a solid basis for clinical decision support systems and automated eye disease screening.</p>