Accurate segmentation of chronic wounds from clinical photographs is essential for objective assessment and treatment planning, yet remains challenging due to irregular wound boundaries, heterogeneous backgrounds, and varying imaging conditions. This article presents EUNet, an attentionenhanced encoder-decoder framework for automatic chronic wound segmentation. EUNet builds on a ResNet50 backbone and integrates squeeze-excitation attention, residual depth-wise separable convolutions, a dilated bottleneck, and a shallow fullresolution skip connection to jointly capture global context and fine boundary details. The model is trained end-to-end using a combined binary cross-entropy and Dice loss, mixed-precision optimization, and a warmup cosine learning-rate schedule. Experiments conducted on a composite dataset of 2,760 wound images demonstrate that EUNet achieves a Dice score of 89.23% and an IoU of 8 0. 7 9%, outperforming multiple backbone-based U-Net variants. Ablation and multi-seed evaluations confirm the robustness and effectiveness of the proposed architectural components.