Introduction
Laparoscopic surgeries rely heavily on clear, real-time video feeds. However, tools like electrocautery and lasers burn tissue, generating surgical smoke that degrades visibility and hinders clinical decision-making. While deep learning has produced various image and video "dehazing" methods, they generally treat smoke as a uniform, static degradation.
In reality, surgical smoke has distinct spatio-temporal dynamics. It begins as a localized, directional plume (Diffusion Smoke) and later disperses into a global, directionless haze (Ambient Smoke). Existing methods fail to account for these different physical behaviors, often leaving residual haze or destroying underlying tissue textures.
The STANet Architecture
The authors propose the Smoke-Type-Aware Laparoscopic Video Desmoking Network (STANet), an end-to-end architecture divided into three main sub-networks:
- Smoky Feature Perception: Extracts multi-scale spatial features using a non-rigid trajectory attention module designed to capture the highly deformable nature of smoke.
- Smoke Mask Segmentation: Predicts both the location and the type of smoke. To handle regions where old Ambient Smoke mixes with new Diffusion Smoke, a Coarse-to-Fine Disentanglement Module (C2FDM) uses cross-attention to separate entangled smoke features.
- Smokeless Video Reconstruction: A dual-branch decoder tailored to smoke physics. For Diffusion Smoke, it uses deformable convolutions to track features along directional paths. For Ambient Smoke, it employs adaptive dilated convolutions to capture widespread, non-uniform global distributions.
The STSVD Dataset
Because no existing dataset labels smoke by its motion type, the authors introduced the STSVD (Smoke-Type-Specific Video Desmoking) dataset. This benchmark features 120 paired clean/smoky videos across 28 surgery types, providing pixel-perfect annotations for Diffusion, Ambient, and entangled smoke. This allows the network to learn the specific inductive biases required for each smoke type.
Key Results and Impact
STANet achieves state-of-the-art performance, outperforming 14 existing dehazing and desmoking methods. On the synthetic STSVD dataset, it achieved a PSNR of 33.53 dB, significantly higher than the best existing desmoking method (SelfSVD at 29.51 dB).
Beyond visual restoration, STANet improves downstream surgical AI tasks. Using STANet to pre-process video improved polyp detection (+0.0122 DSC) and surgical instrument segmentation (+0.587 IoU) compared to competing desmoking techniques. This demonstrates that physics-informed desmoking is critical for both human surgeons and computer-aided diagnostic systems.
Summary updated