Avaliação do impacto da forma, textura e campo receptivo na segmentação de vasos sanguíneos

• Ano de defesa: 2025
• Autor(a) principal: Silva, Weslley dos Santos
• Orientador(a): Comin, Cesar Henrique
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC
• Palavras-chave em Português: Segmentação de vasos sanguíneos; Interpretabilidade; Redes neurais; Campo receptivo
• Palavras-chave em Inglês: Blood vessel segmentation; Interpretability; Neural networks; Receptive field
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO
• Link de acesso: https://hdl.handle.net/20.500.14289/23404

Blood vessel segmentation in medical images is a crucial step for the diagnosis of a wide range of pathologies, but it presents unique challenges due to the complexity and variability of the structures involved. Moreover, the lack of interpretability of Convolutional Neural Networks (CNNs) represents a significant barrier to clinical adoption, as it complicates model debugging, limits design improvements, and can reduce diagnostic confidence. In this study, we conducted systematic experiments to separately evaluate the contributions of shape, texture, and receptive field to the performance of CNNs in blood vessel segmentation. First, to assess the role of texture and pixel intensity, these features were selectively removed from image patches using pixel shuffling and normalization. Next, to isolate the influence of shape, segmentation models were trained using only the vessels’ outer contours or centerlines as input, thereby removing internal texture information. Finally, we investigated the amount of context required for robust segmentation by systematically varying the network’s theoretical receptive field. We found that, although pixel intensity is more relevant than vessel texture, CNNs can still maintain high performance even in the absence of both features. In addition, the networks are not able to easily extrapolate the full vessel shape when provided only with contours or centerlines as input. Finally, we observed that the effective receptive field used by the networks is approximately 20 pixels in the datasets employed, although global information can lead to a slight improvement in accuracy for fundus images.