Related Experiment Video
Updated: Aug 19, 2025
Author Spotlight: Automated Deep Brain Stimulation for Parkinson's Disease - Exploring the Possibilities and Challenges of Home Monitoring
Published on: July 14, 2023
BronchoPose: an analysis of data and model configuration for vision-based bronchoscopy pose estimation
Juan Borrego-Carazo1, Carles Sanchez2, David Castells-Rufas3
1Computer Vision Center, Universitat Autònoma de Barcelona, Cerdanyola del Vallès 08193, Spain; Department of Microelectronics & Electronic Systems, Universitat Autònoma de Barcelona, Cerdanyola del Vallès 08193, Spain.
This study introduces a new synthetic dataset and deep learning models for improved bronchoscopy navigation and tracking. These advances enhance accuracy and reduce memory usage, benefiting future research.
Area of Science:
- Computer Vision
- Medical Imaging
- Artificial Intelligence
Background:
- Advances in neural networks and temporal image processing offer new possibilities for vision-based bronchoscopy tracking.
- Progress has been limited by a lack of standardized experimental data conditions for method comparison.
- Deep learning for temporal structures in bronchoscopy navigation remains an underexplored area.
Purpose of the Study:
- To address the lack of comparative data by introducing a novel synthetic dataset for bronchoscopy navigation and tracking.
- To investigate deep learning architectures for learning temporal information in bronchoscopy, exploring different levels of personalization.
- To provide new insights and results for improving bronchoscopy navigation.
Main Methods:
- Utilized a novel synthetic dataset for bronchoscopy navigation and tracking.
- Explored deep learning temporal information architectures, including Recurrent Neural Networks and 3D convolutions, with an EfficientNet-B0 backbone and ShuffleNet blocks.
- Investigated various loss functions for rotation tracking and population modeling schemes (personalized vs. population).
Main Results:
- Temporal information architectures significantly improved both position and angle estimation accuracy.
- Personalized models demonstrated benefits over population-based schemes, and appropriate loss metrics enhanced results.
- Achieved superior performance compared to a state-of-the-art model, with 12.2% and 18.7% improvement in position and rotation, respectively, and a 67.6% reduction in memory consumption.
Conclusions:
- Advances in temporal information architectures, loss configuration, and population schemes enhance the state-of-the-art in bronchoscopy analysis.
- The introduction of the first synthetic dataset provides a crucial benchmark for fair method comparison, fostering further research in bronchoscopy.
