Optical flow estimation is a crucial task in computer vision, aiming to infer motion information in scenes by analyzing pixel movements in video sequences. Traditional methods rely on photometric consistency assumptions and motion smoothness constraints but often struggle with large motions, complex scenes, or occlusions. Deep learning-based approaches have significantly advanced the field, with notable contributions from FlowNet, FlowNet2.0, PWC-Net, and RAFT, enhancing accuracy and robustness. However, challenges remain, including robustness in complex environments and accurate estimation of large displacements. To address these, we propose MSI-Net, a deep optical flow estimation network that integrates multi-scale feature extraction, deformable convolution, and a layer-by-layer refinement strategy. MSI-Net enhances robustness and accuracy, especially when handling large motions and occlusions. Experimental results on synthetic (Sintel) and real-world (KITTI) datasets demonstrate the effectiveness of our approach, which achieves comparable or superior accuracy to state-of-the-art methods while maintaining high efficiency.