Reconstruct Dynamic Soft-Tissue With Stereo Endoscope Based on a Single-Layer Network.

Area of Science:

• Computer Vision
• Medical Robotics
• Surgical Navigation

Background:

• Accurate 3D reconstruction of deformable soft tissues is crucial for minimally invasive surgery.
• Existing methods face challenges in dynamic surgical environments using stereo endoscopic images.
• Integrating geometric deformable models with neural networks offers a promising approach.

Purpose of the Study:

• To propose a novel self-supervised stereo reconstruction framework for deformable soft tissues.
• To enhance the accuracy and efficiency of 3D surface reconstruction in real-time surgical scenarios.
• To bridge traditional geometric deformable models with modern neural network architectures.

Main Methods:

• Investigated the equivalence between the thin plate spline (TPS) model and neural networks (fully-connected/convolutional).
• Developed a self-supervised framework with alternating training of two TPS-equivalent networks.
• Learned disparity priors from historical stereo frames to create an optimized disparity basis for accurate estimation.

Main Results:

• The proposed framework demonstrated improved accuracy in estimating disparity maps for subsequent frames.
• The method maintained computational efficiency and robustness in dynamic surgical environments.
• Validation was performed using stereo-endoscopic videos from da Vinci surgical robots.

Conclusions:

• The self-supervised framework effectively addresses the challenges of 3D soft-tissue reconstruction in minimally invasive surgery.
• The integration of geometric models and neural networks provides a robust and efficient solution.
• The learned disparity priors significantly enhance the accuracy of 3D surface reconstruction from stereo endoscopic data.