Temporal Memory Enhancement for Semantic Segmentation in Surgical Video

Segmenting critical anatomical structures in surgical videos can enhance precision and patient safety by alerting surgeons to potential complications. While current methods that store features from past frames have advanced the performance in video segmentation, their reliance on a fixed-range local memory often fails to capture complex temporal contexts of surgical scenes. Specifically, the memory could fill with redundant features or omit informative frames due to the non-uniform rate of operations by the surgeons. Besides, the image features in the same phase of the surgery share similar patterns, while local memory could not capture such long-term relationships. Therefore, we propose a memory enhancement method to enrich the local temporal context and incorporate global phase context for surgical video semantic segmentation. Concretely, we improve the local memory with a feature selection module based on Determinantal Point Process (DPP) to choose past features that are diverse and relevant to the current feature. Besides, we introduce a global memory to store the common patterns of frames within each phase based on the conditional variational autoencoder with a mixture of Gaussian priors (CVAE-MoG). Experiments on endoscopic submucosal dissection (ESD) and laparoscopic cholecystectomy (LC) video segmentation demonstrate that our method achieves superior performance over existing methods.