A Self-Supervised Network-Based Smoke Removal and Depth Estimation for Monocular Endoscopic Videos.

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Technology

Background:

• Accurate depth estimation in monocular laparoscopic videos is crucial for minimally invasive surgery but is hindered by smoke.
• Existing methods often require manual labels or patient-specific data, limiting their applicability.

Purpose of the Study:

• To develop a label-free, self-supervised method for monocular endoscopic depth estimation that effectively removes smoke.
• To improve the accuracy and real-time performance of depth estimation in challenging surgical environments.

Main Methods:

• A two-step approach: de-endoscopic smoke removal using a cyclic GAN (DS-cGAN) with a generator network (SGEM, RDBM, RUCM), followed by depth estimation using a high-resolution residual U-Net (HRR-UNet).
• The HRR-UNet incorporates a DepthNet and two PoseNets, utilizing adjacent frames for camera self-motion estimation.
• The method is self-supervised, requiring no manual labeling or CT scans.

Main Results:

• The proposed method effectively removes smoke while preserving critical surgical site details like blood vessels, contours, and textures.
• Achieved accurate depth information in real surgical scenes, outperforming state-of-the-art methods.
• Demonstrated real-time performance with a frame rate of 94.45fps.

Conclusions:

• The developed self-supervised, smoke-removal depth estimation method offers a promising solution for enhancing surgical navigation in minimally invasive procedures.
• Its label-free nature and real-time capability make it suitable for clinical applications.