BronchOpt : Vision-Based Pose Optimization with Fine-Tuned Foundation Models for Accurate Bronchoscopy Navigation

BronchOpt represents a significant advancement in bronchoscopy navigation, tackling the persistent challenges of intra-operative localization due to factors like respiratory motion and anatomical variability. Traditional vision-based methods have struggled with generalization across different patients, often resulting in alignment errors. The introduction of BronchOpt not only provides a robust framework for 2D-3D registration between endoscopic views and pre-operative CT anatomy but also establishes a new synthetic benchmark dataset aimed at standardizing evaluations in this field. The framework's performance is noteworthy, achieving an average translational error of just 2.65 mm and a rotational error of 0.19 rad, demonstrating its potential for improving surgical outcomes. Qualitative results on real patient data further validate the model's effectiveness, marking a promising step forward in enhancing the precision of bronchoscopy procedures.