SPICE: Leveraging Soft Probabilistic Causal Intervention for Breast Ultrasound Tumor Segmentation

Abstract

Accurate segmentation of breast tumors in ultrasound images is essential for clinical diagnosis and treatment planning. However, ultrasound imaging is inherently affected by speckle noise, low contrast, and scanner-specific artifacts, which can introduce spurious correlations and limit the reliability of conventional deep learning models. To address these challenges, we propose SPICE (Soft Probabilistic Intervention for Causal sEgmentation), a causal intervention framework that explicitly disentangles anatomy-consistent causal features from unstable imaging confounders in breast ultrasound images. SPICE adopts probabilistic causal-confounding disentanglement with dual-branch supervision to structurally separate lesion-related and acquisition-dependent representations. Additionally, causal-confounding contrastive learning and multiscale causal consistency regularization are incorporated to enhance feature discriminability and stability. Experiments on five public datasets demonstrate that SPICE achieves superior segmentation performance with the Dice similarity coefficients of 88.42% on the internal validation cohort and 89.49% and 84.14% on two external cohorts, outperforming other state-of-the-art methods. SPICE also provides explicit causal and confounding outputs, enabling interpretable and uncertainty-aware predictions. These results indicate that structured causal intervention enhances both segmentation accuracy and reliability in breast ultrasound imaging.