Joint color-spatial iterative interaction and metric-based motion filtering for unsupervised polyp segmentation in endoscopic videos

Accurate segmentation of polyps in endoscopic videos is essential for early detection and improving patient survival. However, existing polyp segmentation methods primarily rely on fully supervised or semi-supervised learning, both of which require large amounts of manually labeled data for model training, leading to extremely high annotation costs. To address this issue, we propose a novel Joint Color-Spatial Iterative Interaction and Metric-Based Motion Filtering Unsupervised Learning (JCM-SSL), which enables effective segmentation of polyps in endoscopic videos without the need for labeled data during training. Importantly, JCM-SSL is the first to integrate color-spatial iterative interactions with a low-dimensional motion embedding approach, allowing for synergistic extraction and refinement of both morphological and motion features. In addition, JCM-SSL incorporates an innovative time-gating mechanism that adaptively adjusts the most valuable contributions during feature propagation, ensuring the effective temporal alignment and fusion of polyp morphology and motion patterns. Evaluations on the PolyVid-SEG, LDPolypVideo, and CVC-ColonDB datasets show that JCM-SSL achieves at least 3.3% higher Dice scores compared with existing self-supervised methods that do not require labeled data, while remaining competitive with annotation-dependent supervised and domain-specific self-supervised approaches.