Multimodal magnetic resonance imaging (MRI) often faces challenges such as insufficient model generalization ability and high misdiagnosis rates in intelligent neuroimaging diagnosis. Existing methods, including single-path classification networks such as residual networks and EfficientNet, and segmentation-oriented workflow models such as U-Net, have the following limitations when processing multi-sequence MRI data: ① single-path structures have difficulty in effectively decoupling modality-specific features; ② conventional convolutions lack adaptive enhancement in the channel and spatial dimensions. Therefore, this paper proposes a dual-path residual channel-spatial attention network (DRCSA-Net). The model learns complementary representations through a parallel dual-branch structure (channel enhancement and spatial enhancement), introduces a squeeze-and-excitation module and a channel-spatial attention module to achieve channel recalibration and spatial attention focusing, and finally integrates information and completes classification through lightweight fusion convolution and fully connected layers. To comprehensively evaluate the proposed model, five-fold cross-validation, single-/dual-path comparison, ablation experiments, and robustness analysis were conducted on four public datasets. The experimental results show that DRCSA-Net achieves high accuracy on all four public datasets and demonstrates good effectiveness and stability in the single-/dual-path comparison, ablation experiments, and robustness analysis, providing a structurally clear and stable implementation scheme for brain tumor MRI classification.