ACD-Desmoke: Adaptive Constrained Diffusion for Generalizable Zero-Shot Surgical Smoke Removal in Laparoscopy
Year: 2026
Type: conference-paper
Abstract: Laparoscopic surgery is frequently hindered by surgical smoke, which degrades visibility and interferes with surgeons' operations. While existing desmoking methods attempt to address this issue, their reliance on i) expensive synthetic datasets or ii) intricate unpaired training architectures fundamentally limits real-world clinical applicability. To address these challenges, we attempt to establish a paradigm shift and propose ACD-Desmoke (Adaptive Constrained Diffusion Desmoking), a zero-shot, training-free framework. Specifically, we leverage pre-trained diffusion models to implement adaptive constrained diffusion steered by universal priors, which navigates the sampling process toward a reliable solution space, fundamentally bypassing the requirements for task-specific training data and explicit degradation modeling. Furthermore, by introducing adaptive weight allocation and timestep control mechanisms, ACD-Desmoke effectively balances multiple optimization objectives to ensure high-quality image restoration. Extensive experiments demonstrate its superior generalizability, achieving the best training-free performance with a BRISQUE score of 9.543 and PAQ-2-PIQ score of 73.520 on the CholecT50-smoky dataset. This offers a practical, zero-shot solution for surgical smoke removal without the need for task-specific training.
Language: English
Topic: Image Enhancement Techniques
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7172016337
