The accurate prediction of the left ventricular ejection fraction (LVEF) in echocardiograms is limited by low image resolution, artifact interference, and the complexity of cardiac motion. These challenges lead to significant deficiencies in most existing methods in terms of local spatio-temporal dependency modeling and fine-grained motion capture. To address these issues, this paper proposes an Optical Flow-Augmented Dual-Stream Network (OFDS). The innovation of this model lies in the construction of an optical flow feature enhancement branch to compensate for the limitations of the global spatio-temporal feature extraction method based on Uniformer. Specifically, the optical flow enhancement branch optimizes the local motion representation ability through the following techniques: 1) Segmentation Enhanced Feature Extraction Module: by combining semantic segmentation and optical flow calculation, it enables the extraction of targeted optical flow features of the left ventricle, effectively isolating the motion interference from non-target regions such as other chambers; 2) Optical Flow Correction Module: by introducing the irrotational condition constraint and the strain rate consistency constraint, it adaptively corrects the optical flow features in the ultrasound modality based on the reconstruction task, overcoming the adaptability limitations of traditional optical flow algorithms for low-quality ultrasound sequences; 3) Adaptive Gating Fusion Module: it dynamically integrates the global spatio-temporal features extracted by Uniformer with the corrected local optical flow features, balancing the global structure and local motion information, and improving the robustness of LVEF calculation. Our model is validated on two datasets and the experimental results demonstrate that OFDS outperforms the state-of-the-art methods.