Accurate multi-organ semantic segmentation in laparoscopic surgery videos is essential for computer-assisted interventions, enabling organ recognition, instrument tracking, and context-aware guidance. However, surgical scenes exhibit several intrinsic challenges: large non-rigid organ deformations, occlusions caused by instruments or blood, rapid viewpoint changes, motion blur, and heterogeneous lightning. These factors undermine the temporal stability and boundary precision of conventional single-frame segmentation networks. To overcome these limitations, this thesis proposes SSTC-Seg, a deformable memory-based multi-scale architecture specifically designed to enforce spatial adaptivity and temporal coherence in minimally invasive surgical environments. SSTC-Seg integrates three complementary components into a unified, end-to-end trainable architecture. First, a deformable multi-scale encoder —combining deformable convolutions, multi-scale feature extraction, and lightweight self-attention— adapts receptive fields to non-rigid anatomical structures while capturing both fine-grained and global contextual cues.