In this paper, we propose a novel multi-scale gradient computation framework with shape-dependent analysis. Within localized neighborhood windows, shape-dependent gradients are estimated using predefined structural templates. Directional gradients are then obtained by averaging responses across eight orientations around each central pixel at multiple scales. To suppress noise while preserving fine structures such as edges, textures, and spot-like features, gradients derived from three distinct templates are combined through an adaptive aggregation scheme. The resulting multi-scale gradient maps are subsequently fused into a unified representation that highlights structural details of varying sizes. This fused gradient is incorporated into an enhanced active contour model, which effectively leverages multi-scale information to guide contour evolution with higher precision. Experimental results on benchmark datasets demonstrate that the proposed method achieves superior segmentation accuracy and robustness compared to existing single-scale approaches, significantly reducing contour artifacts and improving boundary localization.