<p>Deep neural network-based path generation has demonstrated strong capability in planning static obstacle avoidance maneuvers for autonomous vehicles. However, such learned paths may not always guarantee physical feasibility due to the lack of explicit constraint handling. To overcome this issue, we propose a refinement method based on a Model Predictive Control (MPC) which adjusts the neural network-based path generation result, specifically targeting moderate-speed environments. The MPC problem minimizes a cost function that considers path smoothness, collision safety, and similarity to the neural network-based path while satisfying vehicle kinematics and safety constraints. In evaluating safety within the cost function and constraints, nearby obstacles around the neural network–based path are considered to efficiently monitor potential collisions. Through these processes, the refined path not only preserves the advantages of neural network but also ensures feasibility and safety in execution. Simulation studies show that the proposed MPC significantly improves collision safety compared to solely using the neural network-based path. Moreover, the neural network-based path is shown to effectively guide the MPC solving process with low computational effort by providing a high-quality initial guess and reducing collision constraints.</p>