Megha Kalia, PhD

(She/Her/Hers)

Rank

Fellow or Postdoc

Department

Radiology

Authors

Megha Kalia, Franklin King, Nobuhiko Hata

Principal Investigator

Nobuhiko Hata

Twitter / Website

Categories

Self-Supervised Trackerless Navigation for Bronchoscopy with CT Geometry Prior

Abstract

In bronchoscopy, traditional methods for aligning bronchoscope camera images with preoperative CT scans rely on electromagnetic (EM) trackers, which can become inaccurate due to lung motion and airway deformation during breathing. To address this, we propose a self-supervised pose estimation technique for CT-camera registration. Previous vision-based tracking systems, which estimate 3D scene depth for 3D-3D registration, struggle with visual similarities between the camera view and the CT model. Our method directly estimates the bronchoscope camera pose using a neural network that takes two camera images as input to output the relative pose. We further improve performance by incorporating airway centerline information from the CT scan. Compared to depth-only methods, our approach significantly reduces trajectory errors, with mean root mean square error (RMSE) in the lower right, lower left, upper right, and upper left lung lobes being 20.6 ± 6.7, 15.4 ± 7.2, 14.7 ± 9.9, and 15.5 ± 4.4 mm, respectively, versus 30.5 ± 23, 48.7 ± 24.5, 8.9 ± 3.3, and 35.3 ± 16.2 mm for depth-only methods. In conclusion, our tracker-less pose estimation method provides a more robust and accurate solution for camera-to-CT registration without the need for EM trackers or labeled data.