ResearchPod Summary
In robot-assisted minimally invasive surgery (RAMIS), 3D reconstruction is essential for surgical guidance. However, existing state-of-the-art methods, such as 3D Gaussian Splatting, often rely on offline processing or accurate camera trajectory priors (e.g., from robotic kinematics). These priors are frequently noisy or unavailable in clinical settings. The authors ask: can we perform accurate, online, deformable 3D reconstruction directly from surgical video without relying on precise external pose priors?
The authors introduce Track2Map, a SLAM-based pipeline that treats 3D reconstruction as a joint optimization problem. To handle the inherent ambiguity of non-rigid tissue deformation, the system employs two primary strategies:
Track2Map demonstrates superior performance on the StereoMIS dataset compared to existing SLAM and non-SLAM baselines. It maintains high reconstruction quality (PSNR/SSIM) regardless of whether pose priors are clean, noisy, or entirely absent. The motion-aware gating and deformation anchoring are shown to be critical, as ablations reveal that removing these components leads to significant performance degradation, particularly under heavy pose noise.
By decoupling the reliance on robotic kinematics, Track2Map increases the clinical applicability of 3D reconstruction in surgery. It provides a robust framework for generating high-fidelity, deformable 3D models in real-world scenarios where camera tracking is unreliable or unavailable, facilitating better augmented reality overlays and surgical guidance.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.