Nonhomogeneous smoke mitigation in laparoscopic images using a hybrid physical-neural model.

Surgical smoke generated during electrocautery in minimally invasive surgery introduces nonhomogeneous optical scattering that compromises visibility and clinical interpretability of laparoscopic images. This work proposes a hybrid physics- and learning-based method for surgical smoke mitigation combining a convolutional neural network-vision transformer (CNN-ViT) architecture, which jointly estimates the spatially varying transmission map and ambient illumination, with a physics-based restoration model derived from atmospheric optics. Training data are generated through a Navier-Stokes fluid dynamics simulation of surgical smoke, producing physically grounded, nonhomogeneous degradation patterns representative of intraoperative conditions. The method was evaluated on different unseen laparoscopic sequences and compared against four state-of-the-art baselines, consistently outperforming all methods in both PSNR and SSIM. Evaluation on real electrocautery smoke demonstrates effective generalization to clinically realistic conditions, recovering anatomical structures and tissue detail without reliance on synthetic smoke degradation.