Vision-Kinematics Interaction for Robotic-Assisted Bronchoscopy Navigation.

Area of Science:

• Medical Robotics
• Computer Vision
• Pulmonary Medicine

Background:

• Minimally invasive endobronchial interventions are crucial for treating lung diseases.
• Vision-based localization for bronchoscopy is preferred but vulnerable to visual changes.
• Static pre-operative planning hinders real-time anatomical mapping in learning-based methods.

Purpose of the Study:

• To develop a robust navigation framework for monocular bronchoscopic videos.
• To address visual domain-imbalance and enhance camera pose estimation accuracy.
• To improve the clinical applicability of vision-based localization in endobronchial procedures.

Main Methods:

• Proposed a Vision Kinematic Interaction (VKI) navigation framework.
• Introduced a Visual Similarity Network (VSN) for domain-invariant and domain-specific feature extraction.
• Developed a Kinematic Refinement Network (KRN) for progressive camera pose refinement.

Main Results:

• The VSN effectively extracts features preserving airway structure despite visual variations.
• The KRN significantly improves the robustness of online camera pose estimation.
• Achieved Absolute Trajectory Errors (ATE) of 8.01 mm on phantom data and 8.62 mm on porcine data.

Conclusions:

• The VKI framework offers a robust solution for camera localization in endobronchial interventions.
• The proposed VSN and KRN enhance navigation accuracy and reliability in challenging visual conditions.
• This approach advances minimally invasive pulmonary disease treatment through improved robotic bronchoscopy.