Hybrid Deep Learning Architecture for Retinal Vessel Segmentation: Integrating Attention U-Net and Resunet on the CHASEDB1 Dataset

The computer-aided diagnosis of ophthalmologic diseases, including diabetic retinopathy, glaucoma, and high blood pressure-related retinopathy, depends on retinal vessel segmentation as a key precondition. This study proposes HybridNet, a new deep learning network that combines spatial attention features of Attention U-Net and residual learning features of ResUNet to perform better and distinguish vessels more accurately. The framework adopts a complete preprocessing pipeline that features Contrast Limited Adaptive Histogram Equalization (CLAHE), Gaussian filtering, and intensity normalization to advance the image quality before segmentation. The dual-branch structure has the advantage of channel-wise feature combining via a 1 x 1 convolution layer, which can extract complementary features of attention-guided spatial-wise and context-wise features retained by the residual module. Optimization of the Model was accomplished through a composite loss that uses Binary Cross-Entropy (α =0.7) and Dice loss (β =0.3), trained using the Adam optimizer and adaptive learning rate scheduling over fifty epochs. Experimental optimization on the CHASEDB1 dataset proves its high performance with a Dice coefficient of 0.884, IoU of 0.797, precision of 0.885, recall of 0.877, and AUC-ROC of 0.918. This outperforms baseline designs by +2.8% Dice score and +1.6% IoU. The Data flow modeling is a Crow Foot Notation Entity-Relationship Diagram (ERD) structure that represents the connections and interactions between the users, training sessions, models, metrics, and visual outputs. The results validate that HybridNet performs better than the sum of its parts, in that fine vessel structures, bifurcations, and crossovers are accurately segmented, and background noise is reduced. The framework is shown to have high feasibility in clinical use with automated ocular diagnosis, with its balanced accuracy and computational efficiency in its fast detection.