关键词: 医学图像分割, 活动轮廓模型, Wasserstein距离, 多尺度统计建模, 几何约束损失函数

Abstract: A multi-scale pyramid adaptive Wasserstein geometric constraints loss (PAW) was proposed to address boundary artifacts, internal holes, and false positive predictions in medical image segmentation. First, a Wasserstein-based regional fitting term was constructed based on the optimal transport theory, and the statistical distribution differences between the predicted regions and ground-truth regions were constrained through multi-scale windows to improve regional consistency. Then, a foreground-background adaptive fallback mechanism was introduced to provide stable statistical constraints using global information when sufficient foreground information was unavailable in local regions, thereby reducing the influence of background noise. Finally, a structure-aware gating mechanism was designed to dynamically adjust contour constraints according to boundary gradient information, which enhanced boundary continuity and stability. Experiments were conducted on three public medical image segmentation datasets, including CAMUS(Cardiac Acquisitions for Multi-structure Ultrasound Segmentation), ISIC2017(International Skin Imaging Collaboration 2017), and ACDC(Automated Cardiac Diagnosis Challenge). Experimental results show that the proposed PAW loss function improves segmentation performance on different backbone networks. On the CAMUS dataset, the mDice of TransUNet with PAW increases from 0.896 to 0.902, with an improvement of 0.6 percentage points, while the mBIoU increases from 0.721 to 0.734 by 1.3 percentage points, and the mHD95 decreases from 14.595 to 14.084. Compared with Topology-Aware Loss, Region-wise Loss, and Robust T-Loss, PAW achieves better overall performance in terms of mIoU, mDice, and mBIoU. The results demonstrate that PAW effectively integrates regional statistical constraints and boundary geometric constraints, providing an effective loss function design strategy for medical image segmentation under low contrast and high noise conditions.

Key words: medical image segmentation, active contour model, Wasserstein distance, multi-scale statistical modeling, geometric-constrained loss function