Smoke Attention Based Laparoscopic Image Desmoking Network with Hybrid Guided Embedding

Abstract

Surgical smoke can obscure the surgeon's line of sight during the laparoscopic procedure, potentially posing risk to the patients. Robot assisted-minimally invasive surgery (MIS) also faces such a challenge, which also hinders its autonomous operation. This paper proposes a desmoking method for laparoscopic surgery smoke removal. As the surgical smoke is hetergeneous, a two-stage network is developed to estimate the smoke distribution and reconstruct the corresponding smoke-free scene with the hybrid embedding guidance, including the estimated smoke mask and the initial image. The network is designed following the concept of the atmospheric scattering model, which makes the network more explainable. To qualitatively and quantitatively evaluate the proposed method, a synthetic laparoscopic dataset is generated and a real laparoscopic dataset is obtained. Experimental results show that the Peak Signal to Noise Ratio (PSNR) of the proposed method is 1.9% higher than that of the DehazeFormer, while the run-time is shorter more than 40.9%. The proposed method is comparable to or even outperforms the state-of-the-arts in desmoking quality and computation cost.