Towards Multiple Instance Learning and Hermann Weyl’s Discrepancy for Robust Image-Guided Bronchoscopic Intervention

Abstract

This paper proposes an advantageous approach that introduces multiple instance learning (MIL) and Hermann Weyl’s discrepancy (HWD) to improve image-guided bronchoscopic intervention. Numerous 2D-3D registration methods used for bronchoscopic navigation suffer from problematic bronchoscopic video images (e.g., bubbles and collision) that easily collapse the registration optimization since these images remain challenging to precisely calculate the similarity between bronchoscopic real images and virtual renderings generated from CT slices, resulting in inaccurate bronchoscopic navigation. To address this limitation, we develop a new navigation framework that employs a MIL-driven image classification strategy to remove problematic frames and then performs a HWD-enhanced 2D-3D registration procedure. We validate our framework on patient data. The experimental results demonstrate that our effective and accurate navigation method outperforms others approaches. In particular, the average navigation accuracy of position and orientation was improved from (6.8, 18.0) to 3.5 mm, 9.4$^{\circ }$).