A Two-Stage Post-Hoc Video Event Detection Strategy on Low-Quality Surgical Videos

Real-time video event detection plays a crucial role in enhancing patient care in healthcare settings. However, processing high-quality surgical videos requires significant computational resources, posing a challenge for healthcare institutions with limited infrastructure. Additionally, surgical videos often contain sensitive patient information, which raises concerns regarding privacy. To address these challenges, we propose a novel post-hoc machine learning module designed to leverage low-quality videos for real-time video analysis, reducing both computational requirements and the risk to patient privacy. Specifically, we introduce a two-level label correction method that incorporates a Markov transition matrix and a sliding window-based anomaly detection approach to identify and correct mis-detected events caused by video quality degradation. Our approach allows for effective use of pre-trained video analysis models, ensuring efficient processing of degraded video data while minimizing the need for additional training resources. We evaluate our method using a publicly available real-world clinical procedure dataset, demonstrating its robustness and effectiveness across diverse clinical settings. Our solution offers a practical framework for real-time video analysis in resource-constrained healthcare environments.