PhySe-RPO: Physics and Semantics Guided Relative Policy Optimization for Diffusion-Based Surgical Smoke Removal

Zining Fang; Cheng Xue; Chunhui Liu; Bin Xu; Ming Chen; Xiaowei Hu

cs.AIMar 24, 2026

PhySe-RPO: Physics and Semantics Guided Relative Policy Optimization for Diffusion-Based Surgical Smoke Removal

Authors: Zining Fang, Cheng Xue, Chunhui Liu, Bin Xu, Ming Chen, Xiaowei Hu

Organizations: School of Computer Science and Engineering, Southeast University · Zhongda Hospital,Southeast University · School of Future Technology, South China University of Technology

Abstract

Surgical smoke severely degrades intraoperative video quality, obscuring anatomical structures and limiting surgical perception. Existing learning-based desmoking approaches rely on scarce paired supervision and deterministic restoration pipelines, making it difficult to perform exploration or reinforcement-driven refinement under real surgical conditions. We propose PhySe-RPO, a diffusion restoration framework optimized through Physics- and Semantics-Guided Relative Policy Optimization. The core idea is to transform deterministic restoration into a stochastic policy, enabling trajectory-level exploration and critic-free updates via group-relative optimization. A physics-guided reward imposes illumination and color consistency, while a visual-concept semantic reward learned from CLIP-based surgical concepts promotes smoke-free and anatomically coherent restoration. Together with a reference-free perceptual constraint, PhySe-RPO produces results that are physically consistent, semantically faithful, and clinically interpretable across synthetic and real robotic surgical datasets, providing a principled route to robust diffusion-based restoration under limited paired supervision.

Figures & tables

Figure 1: Limitations of existing restoration approaches: lack of paired data, restoration produce a deterministic output making reward learning difficult, and lack of desmoke restoration-oriented rewards. Our PhySe-RPO addresses these issues by turning restoration into a stochastic policy optimization problem and using physics- and semantics-guided rewards to learn effectively from unlabeled real surgical videos.

Figure 2: Overview of the PhySe-RPO framework. PhySe-RPO refines the pretrained diffusion model through Group-relative Diffusion Policy Optimization, where multiple stochastic trajectories are sampled and optimized using physics-guided color priors, perceptual quality metrics, and semantic rewards, achieving physically consistent and clinically interpretable surgical smoke removal.

Figure 3: Visual-Concept Integration into Diffusion. (a) Learnable visual concepts are trained via contrastive learning to differentiate “clear” and “smoky” concepts in the semantic space. (b) The learned tokens are integrated into the diffusion backbone through multimodal fusion and temporal adaptation to guide semantically consistent desmoking.

Figure 4: Qualitative comparison on real-world surgical smoke images. Compared with prior desmoking methods, PhySe-RPO produces clearer structures, more natural color restoration, and fewer residual smoke artifacts.

Method	SSEQ↓	MANIQA↑	PI↓	FADE↓	MUSIQ↑	IS↑	NIQE↓
DCP [ 10 ]	24.944	0.295	3.577	0.401	49.385	2.394	5.732
Desmoke_LAP [ 24 ]	32.305	0.178	5.675	0.604	38.420	2.010	6.504
SelfSVD [ 37 ]	11.868	0.253	3.227	0.415	48.422	2.307	4.320
PFAN [ 43 ]	30.876	0.248	3.562	0.356	46.656	2.365	5.163
Dehamer [ 8 ]	36.741	0.134	4.713	0.516	33.995	2.469	6.349
Tap [ 7 ]	16.646	0.259	3.447	0.428	44.094	2.415	5.582

Table 1: Comparison with state-of-the-art desmoking methods on the real-world surgical dataset. PhySe-RPO achieves the best overall performance across no-reference quality metrics.

Method	SSEQ↓	MANIQA↑	PI↓	FADE↓	MUSIQ↑	IS↑	NIQE↓
Baseline	7.149	0.332	4.111	0.420	45.586	2.613	5.813
Baseline+ $R_{RF}$	5.018	0.318	3.288	0.360	49.582	2.730	4.944
Baseline+ $R_{RF}$ + $R_{VC}$	4.771	0.368	3.235	0.246	53.818	2.764	4.752
Baseline+ $R_{RF}$ + $R_{VC}$ + $R_{PG}$	3.443	0.378	3.125	0.216	54.911	2.797	4.608

Table 2: Ablation results of PhySe-RPO on the real-world dataset.

Method	SSEQ↓	MANIQA↑	PI↓	FADE↓	MUSIQ↑	IS↑	NIQE↓
w/o R ${}_{\text{VC}}$	5.264	0.320	3.664	0.278	51.517	2.639	5.157
Text	4.304	0.308	3.347	0.244	53.269	2.773	4.867
Full	3.443	0.378	3.125	0.216	54.911	2.797	4.608

Table 3: Comparison of different semantic reward strategies on the real-world surgical dataset. The proposed Visual-Concept Reward yields the best perceptual quality and lowest distortion.

Figure 5: Reward convergence analysis. Average reward (a) and reward variance (b) under different semantic reward settings. The Full model converges faster with lower variance than Text-Reward and w/o $R_{\text{VC}}$ .

Method	PSNR↑	CIEDE-2000↓
Desmoke_LAP [ 24 ]	17.97	11.70
DLSI [ 25 ]	19.95	7.94
MS-CycleGAN [ 32 ]	19.44	8.67
GSR [ 29 ]	19.74	8.42
DehazeFormer [ 31 ]	20.80	8.27
Fog-Removal [ 13 ]	20.29	9.45

Table 4: Comparison of image desmoking methods on the Public Surgical Paired Dataset.

Method	IoU↑	Dice↑
MedSAM	0.5807	0.6772
MedSAM+ours	0.5879	0.6834
AMNCutter	0.7087	0.8252
AMNCutter+ours	0.7179	0.8347

Table 5: Results of application in the segmentation domain.

Method	PSNR↑	SSIM↑	LPIPS↓	FID↓
DCP [ 10 ]	29.0136	0.8529	0.1758	121.6056
Desmoke_LAP [ 24 ]	30.9505	0.9144	0.0975	81.8183
SelfSVD [ 37 ]	31.2658	0.8780	0.1237	150.4745
PFAN [ 43 ]	30.5614	0.9030	0.1068	111.8937
Dehamer [ 8 ]	31.6271	0.9627	0.0406	67.8821
Tap [ 7 ]	31.5969	0.9622	0.0472	74.9391

Table 6: Comparison of image desmoking methods on the synthetic dataset in cold start.

Figure 6: Qualitative comparison on synthetic images,compared to other methods, PhySe-RPO exhibits better desmoking capabilities in the cold start phase.

Figure 7: Output results for different $\alpha$ values

Length of token	PSNR↑	SSIM↑
10	35.2440	0.9762
25	35.2552	0.9776
50	35.2558	0.9774

Table 8: Hyperparameter study on token length.

Number of G	SSEQ↓	MANIQA↑	PI↓	FADE↓	MUSIQ↑	IS↑	NIQE↓
2	4.036	0.346	3.296	0.279	53.751	2.796	4.777
4	3.443	0.378	3.125	0.216	54.911	2.797	4.608
8	3.517	0.374	3.235	0.232	55.125	2.806	4.657

Table 9: Hyperparameter study on number of groups

G

Explore similar work

Seeing Through Smoke: Surgical Desmoking for Improved Visual Perception
Mar 26, 2026Jingpei Lu, Fengyi Jiang, Xiaorui Zhang +2Medical Imaging Physics-Guided Image Restoration
Intuitive Surgical, Sunnyvale, CA 94086, USA
SmokeGS-R: Physics-Guided Pseudo-Clean 3DGS for Real-World Multi-View Smoke Restoration
Apr 7, 2026Xueming Fu, Lixia Han3D Reconstruction Multi-View 3D Reconstruction
University of Science and Technology of China · Nanjing University of Aeronautics and Astronautics
SurgiATM: A Physics-Guided Plug-and-Play Model for Deep Learning-Based Smoke Removal in Laparoscopic Surgery
Nov 7, 2025Mingyu Sheng, Jianan Fan, Dongnan Liu +3Physics-Guided Image Restoration Image Restoration
School of Computer Science, The University of Sydney, Sydney, NSW, Australia · Institute of Medical Robotics, School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China · Department of Radiology, Brigham and Women’s Hospital, and Harvard Medical School, Boston, MA, United States

Mar 26, 2026J/K move · Enter open · S save

Seeing Through Smoke: Surgical Desmoking for Improved Visual Perception

Jingpei Lu, Fengyi Jiang, Xiaorui Zhang, Lingbo Jin+1

Intuitive Surgical, Sunnyvale, CA 94086, USA

Medical Imaging Physics-Guided Image Restoration Surgical Robotics

Open paper AskPDF arXiv

Minimally invasive and robot-assisted surgery relies heavily on endoscopic imaging, yet surgical smoke produced by electrocautery and vessel-sealing instruments can severely degrade visual perception and hinder vision-based functionalities. We present a transformer-based surgical desmoking model with a physics-inspired desmoking head that jointly predicts smoke-free image and corresponding smoke map. To address the scarcity of paired smoky-to-smoke-free training data, we develop a synthetic data generation pipeline that blends artificial smoke patterns with real endoscopic images, yielding over 80,000 paired samples for supervised training. We further curate, to our knowledge, the largest paired surgical smoke dataset to date, comprising 5,817 image pairs captured with the da Vinci robotic surgical system, enabling benchmarking on high-resolution endoscopic images. Extensive experiments on both a public benchmark and our dataset demonstrate state-of-the-art performance in image reconstruction compared to existing dehazing and desmoking approaches. We also assess the impact of desmoking on downstream stereo depth estimation and instrument segmentation, highlighting both the potential benefits and current limitations of digital smoke removal methods.

Abstract

Explore similar work

Seeing Through Smoke: Surgical Desmoking for Improved Visual Perception

SmokeGS-R: Physics-Guided Pseudo-Clean 3DGS for Real-World Multi-View Smoke Restoration

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