Cavir: confidence-aware vision-based registration for image-guided surgery.

Abstract

<h4>Purpose</h4>Surgical navigation depends on precise image-to-patient registration, yet line-of-sight issues on the tracker, anatomy deformation, and occlusion from instruments and the surrounding environment degrade accuracy and lead to failing scenarios. We target a tracker-free, robust, and explainable vision-based framework for complex operative scenes.<h4>Methods</h4>We fuse accurate depth estimation with dense 2D point tracking to recover temporally consistent 3D surface motion from surgical video, continuously updating image-to-patient registration. We also present the registration error detection module where we evaluate tracking quality, motion consistency, and 3D motion agreement to rapidly detect and attribute errors (e.g., occlusion, or drift). If an error is detected, the module also disables the mesh overlay.<h4>Results</h4>Across three phantom scenarios (tool-occluded, layer-occluded, and deformed), our method improves target registration accuracy and robustness compared to the traditional method (ICP) and the learning-based baseline (PREDATOR), particularly under partial occlusion and moderate non-rigid manipulation. On the live-porcine sequence, our proposed method delivers higher accuracy and approximately 20 × faster runtime than baselines, indicating feasibility for a real-time application. Beyond accuracy, our approach enables early detection of failure modes, guiding appropriate user intervention.<h4>Conclusions</h4>Our proposed method, coupled with explainable error detection, maintains registration while reducing reliance on external trackers, contributing to the ongoing effort to build trustworthy, resilient image-guided surgery.