Tumor segmentation in ultrasound images plays a critical role in the screening, diagnosis, and prognosis of breast cancer. However, segmentation remains challenging due to factors such as irregular tumor shapes, blurred boundaries, and low contrast between normal and diseased tissues. To address these challenges, we propose the Mamba-guided Causal disentanglement (MaCa) model, which combines causal intervention with Mamba state-space modeling for effective breast tumor segmentation. The MaCa leverages the selective scanning mechanism of the Mamba model to extract global vision features, which are then processed through causal intervention to disentangle causal and confounding features. These features are subsequently reconstructed, and an adversarial paradigm is employed to suppress confounding bias, yielding improved segmentation predictions. Five publicly available breast ultrasound datasets are utilized to evaluate our MaCa model. Extensive experiments demonstrate the state-of-the-art performance of MaCa, which achieves accurate and robust breast tumor segmentation.