Medical image segmentation via shape-dependent anisotropic gradient
Alexander A. Zelensky, Viacheslav Voronin, Evgeny A. Semenishchev, Nikolay Gapon, Marina M. Zhdanova, Ivan Naumov · 2025
Most existing gradient computation methods are sensitive to noise and often struggle to capture fine textural details effectively. In this paper, we propose a novel anisotropic gradient computation technique that is both robust to noise and capable of capturing intricate texture patterns. The method partitions the image into localized neighborhood windows and extracts shape-dependent templates based on predefined structural patterns. Directional gradients are computed via arithmetic averaging in eight orientations around each central pixel. Gradients derived from three distinct structural templates are then aggregated through averaging, which enhances noise suppression while preserving detailed features such as textures, edges, spots, and curves. Furthermore, we integrate the proposed gradient computation into an enhanced active contour model for automated segmentation in medical imaging. Experimental results on benchmark medical image segmentation datasets demonstrate that our adaptive anisotropic gradient reduces contour artifacts and outperforms existing state-of-the-art methods in terms of segmentation quality.