Aiming at the insufficient accuracy in magnetic resonance imaging (MRI) brain tumor segmentation caused by complex tumor shapes and blurred boundaries, we propose a multi-scale convolutional neural network. In the encoding stage, an enhanced multi-branch attention (EMBA) module integrates channel and spatial attention mechanisms: the multi-branch channel attention (MBCA) module adopts hierarchical scaling with multiple scale factors, while the multi-branch spatial attention (MBSA) module leverages diverse receptive fields to improve adaptability to complex structures. During downsampling, a multi-scale module (MSM) replaces single-branch convolution to enhance multi-level feature extraction. In the decoding stage, an adaptive convolution enhancement module (ACEM) substitutes the original single convolution, enabling dynamic fusion of multi-scale features. Experiments on the BraTS 2019 and BraTS 2021 datasets demonstrate that, compared with 3D U-Net, the proposed model achieves Dice coefficient improvements of 9.0%, 8.0%, and 6.6% on BraTS 2019, and 5.6%, 6.6%, and 3.7% on BraTS 2021.