SEmantic Multi-Region Active Contour (SEMAC): uma abordagem semântico-geométrica para segmentação de multirregiões

Abstract

Image segmentation is essential for visual analysis and interpretation, especially in applications that require geometric precision and robustness against variability in shape, texture, contrast, and artifacts. This work proposes the SEMAC, a segmentation method based on active contours whose multi-contour formulation enables the extraction of multiple coherent regions from a single initial contour. This structure allows the contour to divide and adapt dynamically when the energy field identifies multiple nuclei or regions of interest, resulting in consistent multiregional segmentation. The model derives from the classical ACM formulation but redefines its external energy based on a semantic evidence field, which guides the adaptive evolution of contours within a continuous energy domain. This supervised energy replaces the traditional photometric gradient, synthesizing contrasts and structural patterns at multiple scales and ensuring numerical stability even in complex scenarios. The evidence field is obtained from reference image pairs, providing the model with a continuous mapping that reflects the spatial distribution of semantic evidence, making it more informative and robust to photometric variations. The experimental evaluation encompassed three datasets: a synthetic one, aimed at analyzing variations in shape, noise, and density; and two clinical ones, comprising dermatoscopic images of skin lesions and microscopic images of blood cells. The performance metrics (Accuracy, Precision, Sensitivity, Specificity, Intersection over Union (IoU), Dice Similarity Coefficient (Dice) e Matthews Correlation Coefficient (MCC)) were complemented by qualitative analyses of regularity and morphological stability. The results demonstrate that the SEMAC achieves competitive performance compared to classical and state-of-the-art supervised methods, while maintaining consistency and computational efficiency. The model establishes a continuous and multiregional approach inspired by the dynamics of multiple active contours, integrating semantic learning and geometric evolution within an interpretable and stable framework suitable for the segmentation of complex images.