Desenvolvimento de tarefas auxiliares e técnicas de pós-processamento para preservação da topologia em segmentação de vasos sanguíneos

Abstract

Convolutional neural networks (CNNs) have provided important advances in blood vessel segmentation, but ensuring the correct topology and continuity of the segmented vasculature remains a significant challenge. Structural artifacts can compromise subsequent analyses, creating the need for methods that explicitly address vessel connectivity. In this work, we propose the SkelAnchor method, a new auxiliary task designed to improve segmentation connectivity through a post-processing step after CNN-based segmentation. First, we investigate simple graph-based connectivity improvement methods and demonstrate their limitations, showing that they produce discontinuous maps that are difficult for networks to learn. To overcome this, SkelAnchor introduces a smooth and continuous parameterization that encodes the local vessel skeleton distribution using anchor points. This task provides the network with rich topological and geometrical information. Our results show that a model can successfully learn the SkelAnchor task. Although the auxiliary task does not change standard segmentation metrics, we demonstrate through qualitative examples that the learned representation can be successfully used in the post-processing step to repair connectivity problems in the final segmentation. The analysis serves as a proof of concept for the use of skeleton-based auxiliary tasks to enhance the topological quality of blood vessel segmentation.