HG-Mamba: Heuristic-Guided State Space Model for Laparoscopic Image Desmoking

ShiweiWu, XiaoboZhu, Song Zhang, Yu An, Xinyu Zhang, Hui Yan, Jie Tian, Xiyuan Hu, Zhenyu Liu

Keywords: Laparoscopic Surgery, Surgical Smoke Removal, State Space Model, Deep learning

Abstract: Developing smoke removal algorithms for laparoscopic surgery is crucial for enhancing surgical visibility and supporting accurate intraoperative decision-making. Recently, Mamba, a representative state space model (SSM), has shown strong potential in visual tasks by balancing global receptive fields with efficiency. However, its reliance on sequential state transitions limits spatial correlation modeling, and its feed-forward layers lack mechanisms to model frequency features—both of which hinder effective removal of complex surgical smoke. To overcome these limitations, we propose Heuristic-Guided Mamba (HG-Mamba), which extends Mamba by integrating spatial and frequency domain improvements. HG-Mamba comprises two key components: a Heuristic-Guided State Space Model (HG-SSM), which performs input-guided dynamic sampling and flexible state fusion to enable adaptive spatial context modeling; and a Frequency Refine Feed-Forward Network (FR-FFN), which conducts multi-band frequency decomposition and adaptive weighting to enhance frequency-domain representations. By jointly leveraging spatial adaptability and frequency-aware refinement, HG-Mamba serves as an effective backbone for surgical smoke removal. Extensive experiments demonstrate that HG-Mamba outperforms state-of-the-art methods on both synthetic and real-world smoke/smokeless datasets. The code will be publicly released.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 7309