Self-supervised direct pose estimation for vision-based tracking in navigated bronchoscopy

Vision-Based Tracking (VBT) enhances Electromagnetic Tracking (EMT) by addressing Computer CT-to-body divergence. Current VBT methods use depth map-based point clouds to register the bronchoscope in Computer Tomography (CT) scans; however, similarities in airway geometry often lead to registration ambiguities. This paper introduces a Direct Pose Estimation (DPE) method for VBT that bypasses depth maps to localize the bronchoscope in CT scans, validated with human-derived lung phantoms.

We developed a self-supervised DPE (SS-DPE) method for bronchoscopy by introducing a module to mask corrupt pixels and adding pose consistency and long-term consistency losses. We also propose a technique to adjust pose predictions closer to airway centerlines to mitigate drift. The Localization Error and the rate of identifying anatomical airway labels were evaluated using soft phantoms derived from CT of subjects who underwent transbronchial biopsies. SS-DPE was compared to conventional DPE methods using 3cGAN to produce depth maps.

SS-DPE had a statistically significantly lower Mean Localization Error than 3cGAN (p=0.014), with the mean of 20.9±7.5mm and 27.7±12.6mm, respectively. SS-DPE slightly improved the ability to identify anatomical airway labels but without statistical significance over 3cGAN (55.1±20.0% vs. 48.7±26.3%, p=0.505). Pose Consistency Loss function was critical for accurate localization.

Our SS-DPE method, validated on realistic human-derived phantoms, achieved lower Localization Error than the conventional DPE method using 3cGAN, demonstrating its potential for enhancing VBT in navigated bronchoscopy.