Causal networks have attracted significant attention in real-time speech enhancement. However, their performance may be limited by the suboptimal integration of features and rough frequency band modeling. To address these, this paper proposes a causal dual-branch multi-channel network with multi-scale gated feature fusion (CDMGTU-Net). The key innovations are as follows. First, we combine the long short-term memory (LSTM) with convolutional neural networks (CNN) to improve the feature extraction performance, given that they have distinct frequency responses. Second, we introduce the multi-scale gating units, which orchestrate cross-scale flows through hierarchical gating mechanisms, enabling the network to better focus on local patterns of complex spectrograms. Third, we design the multi-scale temporal feature fusion unit, enabling a dynamic fusion of spatial and spectral features. Experiments demonstrate that the CDMGTU-Net significantly outperforms competing causal networks.