Latent Code Predictor for Accelerating Disparity Estimation in Stereo-Endoscopic Surface Reconstruction.

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Technology

Background:

• Disparity estimation from stereo-endoscopic images is crucial for 3D reconstruction in minimally invasive surgery (MIS).
• Challenges include soft tissue deformation, motion blur, and photometric inconsistency in surgical environments.
• Current self-supervised generative networks, like StyleGAN, face high computational latency due to iterative latent optimization, limiting practical use.

Purpose of the Study:

• To accelerate the process of disparity estimation for 3D reconstruction in MIS.
• To reduce computational latency in self-supervised generative networks without sacrificing reconstruction fidelity.
• To enable practical deployment of efficient, self-supervised stereo-endoscopic reconstruction in clinical settings.

Main Methods:

• Proposed a temporal latent prediction method to optimize initial latent vectors for generative networks.
• The framework learns to predict optimized latent vectors, reducing the number of required optimization steps.
• The prediction-guided mechanism maintains the generator's architecture and inference logic, ensuring reconstruction quality.

Main Results:

• Reduced average optimization steps by 16-59% compared to baseline methods.
• Achieved approximately a 2.3× reduction in per-frame latency.
• Demonstrated nearly identical final photometric loss, confirming no compromise in reconstruction quality.
• Experiments conducted on both Phantom and In vivo datasets.

Conclusions:

• The temporal latent prediction method offers a practical solution for accelerating self-supervised stereo-endoscopic reconstruction.
• The approach significantly improves efficiency by reducing latency while maintaining high reconstruction fidelity.
• This advancement facilitates the clinical application of advanced 3D reconstruction techniques in MIS.