SRTD-Net: Decoupling Spatial Reliability from Temporal Dynamics for endoscopic video segmentation under instrument–tissue interactions

Real-time segmentation of surgical instruments and anomalies in endoscopic videos is essential for image-guided surgery, yet remains challenging due to high temporal heterogeneity, frequent occlusions, and complex non-linear instrument–tissue interactions that often lead to temporal inconsistency and visual ambiguity. We present SRTD-Net, a Spatial Reliability and Temporal Dynamics Network that explicitly targets spatiotemporal reliability via a dual-stream design that decouples feature refinement from interaction modeling. The Spatial Reliability Stream performs diffusion-inspired feature refinement using an Energy-Calibrated Wavelet Decomposer and a Diffusion Feature Refinement module, suppressing dynamic noise and restoring corrupted representations. In parallel, the Temporal Modeling Stream integrates Dynamic Kernel Attention to retrieve historical context under occlusion and an Interaction-Gated Temporal Dynamics unit to capture non-linear scene evolution while promoting temporal consistency across frames. To facilitate systematic evaluation under challenging interactions, we further introduce the EndoPolyp-Instrument Dataset (EPID), curated with diverse instrument–tissue and instrument–anomaly interaction scenarios. Extensive experiments on three benchmarks demonstrate that SRTD-Net achieves state-of-the-art performance and shows improved robustness in zero-shot generalization. Code and dataset are available at https://github.com/hedongdong6060/SRTD-Net .