Conceitos essenciais
Novel loss formulation enhances topological correctness in multi-class segmentation.
Resumo
In the realm of medical image segmentation, maintaining topological accuracy is crucial for downstream applications like network analysis and flow modeling. While recent advancements have focused on binary segmentation, multi-class scenarios often face topological errors. This study introduces a general loss function for topologically faithful multi-class segmentation, extending the Betti matching concept. By breaking down the N-class problem into N single-class tasks, training neural networks becomes computationally feasible using 1-parameter persistent homology. The proposed method significantly improves topological correctness in various medical datasets with diverse topological characteristics, including cardiac, cell, artery-vein, and Circle of Willis segmentation.
Estatísticas
"Our loss formulation significantly enhances topological correctness in cardiac, cell, artery-vein, and Circle of Willis segmentation."
"We demonstrate the utility of our method and outperform all baselines."
Citações
"Topological correctness is crucial for many biomedical downstream tasks."
"Our method significantly improves the Betti matching error and the Betti number errors."