Accurate segmentation of tree-like structures, such as airway and vessel trees, is crucial for medical image analysis. However, this task is challenging due to the inherent topological complexity of these structures. Moreover, the immense data annotation cost underlying this complexity often leads to incomplete annotations in practice. To address these challenges, we propose a novel framework called Tree-Skeleton Collaborative Learning (TSCL). By integrating tree and skeleton segmentation with bias correction, TSCL significantly enhances topological completeness. The core idea of this study is to leverage skeleton guidance and dynamic pseudo-label generation to enhance fine-branch detection and continuity. Additionally, we introduce the Expectation-Maximization (EM) algorithm to address annotation bias, identifying and correcting missing branches in incomplete labels. This approach enables TSCL to enhance the topological completeness of the segmentation while also maintaining high-performance accuracy. Experiments on two public datasets (BAS and PARSE) confirm that TSCL outperforms state-of-the-art methods in topological completeness metrics, thereby proving its robustness and reliability for tree-like structure segmentation.
Zhou et al. (Tue,) studied this question.