Evaluating Segmentation Using Betti-1 Topological Metric: Application to Nasal Cavities in the Context of Airflow Simulation

Abstract

Computational fluid dynamics (CFD) simulations offer an objective means to analyze nasal airflow. However, to be patient-specific, their accuracy relies on precise CT-based volume segmentation of the nasal cavity. Existing segmentation methods typically prioritize volumetric accuracy (e.g., Dice coefficient) while often overlooking topological fidelity, which is critical for generating anatomically consistent CFD meshes. To address this limitation, we propose an automatic nasal cavity segmentation framework based on nnUNet, augmented with a topology-based evaluation metric that quantifies differences in the number of tunnels between predicted and reference segmentations. Evaluation on the NasalSeg public dataset shows a Dice score of 0.947, comparable to state-of-the-art results, while improving the segmentation consistency by removing tunnels.