Spatial-Temporal Consistency Enhanced Segmentation for Laparoscopic Surgical videos

Lin Guo, Ewen Rondel, Flavio Esposito, Mohammad Mahmoud

Keywords: Computer Vision, Segmentation, CRF, Machine learning

TL;DR: This paper introduces SSTC-Seg(Surgical Spatial-Temporal Consistency Segmentation), a lightweight yet effective segmentation framework designed to address the challenges of multi-organ segmentation in laparoscopic surgical videos.

Abstract:

Accurate segmentation of laparoscopic surgical videos is essential for enhancing intraoperative guidance and improving patient outcomes. However, this task remains challenging
due to the constrained field of view, visual clutter, frequent occlusions, and inconsistent illumination. To address these challenges, we propose SSTC-Seg (Surgical Spatial-Temporal
Consistency Segmentation), a lightweight deep learning framework for video-based segmentation. It integrates a memory system and a Hierarchical Dense Conditional Random
Field (HD-CRF) with skip connections for spatial details preservation to refine coarse
predictions and model contextual relationships across frames. Evaluated on the Dresden
Surgical Anatomy Dataset (DSAD), SSTC-Seg achieves competitive multi-organ segmentation performance with significantly fewer parameters compared to existing state-of-the-art
methods.

Submission Number: 104