TASTE:Triple-attention with weighted skeletonized Tversky loss for enhancing airway segmentation accuracy

Abstract

Airway segmentation plays a crucial role in medical image processing. However, the accuracy and efficiency of existing segmentation methods still cannot meet the demands of practical applications. This paper proposes a novel airway segmentation method based on 3D UNet, which integrates a triple-attention mechanism and a new loss function based on skeletonization to improve the accuracy of airway segmentation. First, we obtain the multi-scale connectivity features and attention map by constructing a connectivity matrix. Then, by combining this attention map, we introduce spatial and channel attention mechanisms. Additionally, we incorporate an airway skeletonized loss function. This approach effectively address discontinuity issues and class imbalance in airway segmentation tasks, thereby improving the accuracy of airway segmentation. To validate the effectiveness of the method, we conducted a series of experiments on a publicly available dataset. The experimental results demonstrate significant performance improvements compared to the state-of-the-art methods in most metrics, especially in DLR and DBR, reaching 95.8% and 92.5%.