Cable segmentation is a critical component in industrial automation, large-scale equipment inspection, and the emerging field of embodied AI. However, it remains a challenging task due to the diverse morphology of cables (varying thickness, complex twisting), intricate layouts involving self-occlusion and external occlusion, and cluttered backgrounds that make distinction difficult. While recent advancements in Convolutional Neural Networks and Transformers, including the Segment Anything Model, have improved general segmentation, they often struggle with the discontinuity and incompleteness associated with thin, entangled cables. To address these limitations, this paper proposes a novel cable segmentation framework based on Mamba, integrated with a U-Net architecture. We construct a multi-scale feature extraction module to capture global context and local de-tails simultaneously. Furthermore, to mitigate class imbalance and boundary fuzziness, we introduce a composite loss function combining Adaptive Weighted Cross-Entropy, Similarity Loss, and Boundary-Aware Loss. Extensive experiments on the Synthetic segmentation dataset, Electric Wires (C1/C2) datasets, and a real-world cabin cable dataset demonstrate that our method significantly outperforms state-of-the-art baselines in terms of segmentation accuracy and boundary precision, proving its efficacy in complex industrial environments.