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

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Technology

Background:

• Effective surgical smoke removal algorithms for laparoscopic surgery are lacking due to the absence of real-world paired datasets.
• Existing de-smoking methods rely on synthetic data and non-reference metrics, failing to represent in vivo complexities.

Purpose of the Study:

• To introduce a novel paired dataset of real smoky and smoke-free laparoscopic surgical scenes.
• To evaluate current de-smoking algorithms and critically analyze underlying models using this new dataset.

Main Methods:

• A robust motion-tracking technique was employed to compensate for patient movement, enabling reliable pairing of smoky and smoke-free images.
• A dataset of 3000 image pairs was curated from laparoscopic prostatectomy and cholecystectomy recordings.
• Commonly used atmospheric scattering models and the dark channel prior were revisited and analyzed.

Main Results:

• Current de-smoking methods show effectiveness but also limitations when evaluated on the new dataset.
• The traditional atmospheric scattering model with a "gray smoke" assumption introduces significant errors in green and blue channels.
• The dark channel prior correlates strongly with smoke intensity, suggesting its potential as an attention map for deep learning models.

Conclusions:

• The developed paired dataset addresses a critical gap in training and evaluating surgical smoke removal algorithms.
• The analysis provides insights into the limitations of existing models and the potential of the dark channel prior for advanced de-smoking techniques.
• This work facilitates the development of more robust and accurate de-smoking solutions for laparoscopic surgery.