SurgiATM: A Physics-Guided Plug-and-Play Model for Deep Learning-Based Smoke Removal in Laparoscopic Surgery

TL;DR

Surgical smoke removal in laparoscopy is critical for visibility and downstream tasks. The paper introduces SurgiATM, a physics-guided, plug-and-play module with zero trainable weights that blends a denormalized dark-channel-based physics prior with DL predictions to yield a new restoration formula $J_{MoE}(x,c)= I(x,c) - \mathcal{D}(x) \cdot (1 - \rho_{DNN}(x,c))$, eliminating the need to estimate global illumination. This approach improves restoration accuracy and generalization across three public datasets and ten baselines, while requiring minimal integration overhead. The work demonstrates that physics-guided MoE can stabilize desmoking and enhance performance in real-world surgical scenes, with code available at the project URL for reproducibility and adoption in clinical imaging pipelines.

Abstract

During laparoscopic surgery, smoke generated by tissue cauterization can significantly degrade the visual quality of endoscopic frames, increasing the risk of surgical errors and hindering both clinical decision-making and computer-assisted visual analysis. Consequently, removing surgical smoke is critical to ensuring patient safety and maintaining operative efficiency. In this study, we propose the Surgical Atmospheric Model (SurgiATM) for surgical smoke removal. SurgiATM statistically bridges a physics-based atmospheric model and data-driven deep learning models, combining the superior generalizability of the former with the high accuracy of the latter. Furthermore, SurgiATM is designed as a lightweight, plug-and-play module that can be seamlessly integrated into diverse surgical desmoking architectures to enhance their accuracy and stability, better meeting clinical requirements. It introduces only two hyperparameters and no additional trainable weights, preserving the original network architecture with minimal computational and modification overhead. We conduct extensive experiments on three public surgical datasets with ten desmoking methods, involving multiple network architectures and covering diverse procedures, including cholecystectomy, partial nephrectomy, and diaphragm dissection. The results demonstrate that incorporating SurgiATM commonly reduces the restoration errors of existing models and relatively enhances their generalizability, without adding any trainable layers or weights. This highlights the convenience, low cost, effectiveness, and generalizability of the proposed method. The code for SurgiATM is released at https://github.com/MingyuShengSMY/SurgiATM.

Figures (5)

• Figure 1: The left images are from a natural dehazing dataset, O-HAZE Ancuti_2018_O_HAZE; the middle group is from a real-world surgical desmoking benchmark VASST-desmoke Xia_2025_In_Vivo; and the line chart (right) shows the difference in error distributions between surgical desmoking and natural dehazing, computed from the entire dataset. The $\text{1}^\text{st}$ row displays the hazy or smoky images; the $\text{2}^\text{nd}$ row shows the smoke or haze mask estimated from the ground truth; and the $\text{3}^\text{rd}$ row presents the error magnitude map of the DCP prediction.
• Figure 2: The Laplacian parameters (i.e., $\mu$ and $b$) estimated from the specific dataset, and indicator $W^*$ calculated with \ref{['Equ_Method_EMM_MoE_ERR_OPT_SOLUTION']}, are presented across different methods with the Confidence Interval (CI). The upper and lower rows correspond to Cholec80 and VASTT-desmoke, respectively.
• Figure 3: Error comparison in VASST-desmoke. “+” indicates the baseline integrated with our SurgiATM. "GT" represents ground truth. The methods are trained in Cholec80 with synthetic smoke. Orange boxes mark the errors declined by SurgiATM. Overall, “Baseline + SurgiATM” exhibits more accurate restoration.
• Figure 4: Comparison of SurgiATM with baselines on real datasets: Cholec80 (left two columns) and the Hamlyn Dataset (right two columns). Blue boxes highlight the comparatively accurate and stable restoration achieved by SurgiATM.
• Figure 5: Each heatmap illustrates the ablation study for a specific baseline model incorporated with SurgiATM, showing the impact of different hyperparameter settings in terms of the RMSE metric (lower is better). Each heatmap uses an independent color scale.