Visual-inertial odometry (VIO) remains at the forefront of recent advances in computer vision and robotics. Sparse feature-based VIO is a representative and reliable approach that provides accurate pose estimation through robust feature extraction and tracking. Recent advancements in deep learning have demonstrated the potential of dense optical flow, which predicts pixel-wise motion and can replace sparse features for pose recovery or Kalman filter updates as visual measurements. However, its effectiveness depends on network architecture and pre-training, and the impact of these factors on VIO performance still requires further research. To address these challenges, this paper proposes DeepOF-VIO, a filter-based VIO system that integrates deep optical flow with a monocular camera and an inertial measurement unit. The proposed method employs deep learning-based optical flow to extract reliable flows and then integrates them with inertial measurements using an extended Kalman filter. High-quality flow is extracted by enforcing a bidirectional adaptive consistency constraint, combined with visual measurements. Geometry-based pose updates leverage the refined flows, enhancing the estimation accuracy. Experiments on public datasets validate the robustness and accuracy of the proposed method, and the results show competitive performance on pose accuracy compared with other methods. Additionally, quantitative and qualitative analyses examine the influence of different deep optical flow prediction models on VIO performance. © 2025 IEEE.