In Vivo Laparoscopic Image De-smoking Dataset, Evaluation, and Beyond

The development of effective algorithms for removing surgical smoke in laparoscopic surgery has been hindered by the absence of a paired dataset containing real smoky and smoke-free surgical scenes. As a result, existing de-smoking methods have been primarily based on synthetic datasets and non-reference image enhancement metrics, which fail to fully capture the complexity of in vivo surgical scenes. To address this gap, we present a novel paired dataset derived from laparoscopic surgical recordings by identifying video sequences with relatively stationary scenes where smoke emerges. Our approach includes a robust motion-tracking technique that compensates for involuntary patient movements, ensuring reliable pairing of smoky images and their corresponding smoke-free ground truths. From 132 laparoscopic prostatectomy recordings, we curated 41 video sequences, resulting in a dataset of 2000 smoky-to-smoke-free image pairs. From 45 cholecystectomy recordings, we extracted 68 video sequences, resulting in an additional dataset of 1000 image pairs. Using this unique dataset, we evaluated a representative selection of current de-smoking methods, confirming their effectiveness while also highlighting their limitations. Furthermore, we critically revisited the commonly used atmospheric scattering model, atmospheric colour assumptions, and the dark channel prior. Our analysis demonstrated that the traditional atmospheric scattering model with "gray smoke" assumption introduces significant residual errors in the green and blue channels, while the dark channel prior maintains a strong correlation with smoke intensity. These observations suggest that, while less effective for direct smoke separation, the dark channel prior has potential to serve as a useful attention map for deep learning-based de-smoking approaches.