Vascular Skeleton Deformation Evaluation Based on the Metric of Sinkhorn Distance

Vascular malformations, characterized by distortion and narrowing, serve as visible signs of specific vascular disorders. They may underscore the importance of deep learning-assisted vascular segmentation in medical imaging research. However, most existing vascular segmentation methods rely on evaluation metrics based solely on pixel-wise overlap, which, to some extent, masks the true assessment of the model’s ability to capture essential deformations. In this work, we introduce a vascular structure evaluation metric based on Sinkhorn distance called Skeleton Matching Distance(SMD), which focuses on addressing the inappropriate scoring of small blood vessels and skeletal displacement in existing evaluation metrics. Through experiments on simulated scenarios and real datasets, we validated the excellent properties of this metric.