Structural and Contextual Multi-Scale Active Contour (SEMS-AC): um método de segmentação de imagens biomédicas baseado em informações contextuais e estruturais

Abstract

Medical image segmentation is a central challenge in computer-aided diagnosis systems, as it requires morphological precision and robustness in the face of anatomical and pathological variability as well as image acquisition conditions. In this work, we propose the Structural and Contextual Multi-Scale Active Contour (SEMS-AC), a hybrid method based on the integration of contextual and structural information at different scales, which combines contextual activation maps, structural skeletonization, and multiscale forces in a multimodal active contour model. The formulation of the method is organized into two complementary forces: the external energy, called Multiscale Feature Force (MFF), which provides structural support derived from the hierarchical extraction of features; and the internal energy, Structural Activation Force (SAF), constructed from geometric representations of discriminative regions, such as skeletons and central points, capable of guiding the curve in a localized manner consistent with anatomical organization. SEMS-AC was evaluated in two complementary clinical scenarios: lung segmentation in chest computed tomography (CT) images, including healthy individuals and patients with chronic obstructive pulmonary disease (COPD) and pulmonary fibrosis; and segmentation of hemorrhagic lesions resulting from stroke in brain CT images, characterized by high heterogeneity of shape, intensity, and location. Quantitative evaluation included binary classification metrics (Accuracy, Specificity, Precision, Sensitivity, and Matthews Correlation Coefficient) and spatial similarity metrics (Jaccard Index, Dice Coefficient, and Hausdorff Distance), complemented by qualitative analysis. The results showed competitive performance compared to state-of-the-art methods, demonstrating that SEMS-AC is capable of preserving the geometric fidelity of segmented regions and reducing errors associated with diffuse boundaries. Furthermore, the model promotes more transparent and interpretable segmentations, establishing a promising framework for biomedical applications that demand reliability, explainability, and effective support for clinical decision-making.