A 3D reconstruction based on an unsupervised domain adaptive for binocular endoscopy

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Technology

Background:

• Endoscopic image quality is critical for minimally invasive surgery.
• Existing binocular endoscopic images lack true parallax, hindering 3D visualization.
• Surgical smoke degrades image quality and obscures surgical fields.

Purpose of the Study:

• To develop an unsupervised adaptive neural network for enhancing binocular endoscopic images.
• To address the lack of parallax in endoscopic images and improve 3D display.
• To improve surgical visualization by removing smoke and reconstructing 3D tissue structures.

Main Methods:

• Proposed an unsupervised adaptive neural network integrating smoke removal, depth estimation, and 3D display.
• Utilized U-Net fused by Laplacian pyramid for feature extraction and Convolutional Block Attention Module for parameter optimization.
• Employed a self-supervised training approach using disparity transformation for virtual right-eye image generation.

Main Results:

• Successfully removed simulated surgical smoke (fog) from endoscopic images.
• Effectively reconstructed 3D images of tissue structures from binocular endoscope data.
• Preserved crucial details like contours, edges, and blood vessel textures in medical images.
• Demonstrated significant improvements in various performance indicators compared to existing methods.

Conclusions:

• The proposed method enhances binocular endoscopic image quality by removing smoke and reconstructing accurate 3D views.
• The technique effectively preserves fine details, crucial for surgical precision.
• The developed unsupervised adaptive neural network shows strong potential for clinical application in minimally invasive surgery.