Breast tumor segmentation using dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) holds great potential for clinical application. In particular, pre-/post-contrast images and subtraction images emphasize complementary aspects of tumor characteristics, with the former capturing global information such as location and size, and the latter highlighting local details such as boundary delineation. Motivated by this observation, we propose a novel Dual-Encoder Hybrid CNN-Mamba Network (DEHN), which leverages the complementary features of pre- and post-contrast images and their associated subtraction counterparts. The encoder consists of two parallel branches: a Local-Focused Encoder and a Global-Focused Encoder. The Local-Focused Encoder, constructed with convolutional blocks, extracts detailed local features from both pre- and post-contrast images. In parallel, the Global-Focused Encoder integrates convolutional layers with Mamba blocks to capture high-level global contextual features from the subtraction images. The decoder processes the fused local and global features to generate the final tumor segmentation. We evaluated the proposed DEHN on 1,606 cases from two public DCE-MRI datasets (Yunnan and MAMA-MIA). The proposed DEHN consistently outperformed baseline and seven state-of-the-art models across key metrics, including Dice coefficient, precision, and recall. Specifically, DEHN achieved Dice scores of 81.52% on the Yunnan dataset and 75.93% on the MAMA-MIA dataset.