Airway label prediction in video bronchoscopy: capturing temporal dependencies utilizing anatomical knowledge.

Area of Science:

• Medical Imaging
• Computer Vision
• Robotics

Background:

• Bronchoscopy navigation is crucial for lung interventions but often relies on electromagnetic tracking and patient-specific CT scans.
• These requirements limit applications in settings like intensive care units due to workflow and hardware constraints.

Purpose of the Study:

• To develop a bronchoscope tracking system using only video data, removing the need for electromagnetic tracking and CT scans.
• To enable navigation guidance for lung interventions without altering clinical workflows or requiring additional hardware.

Main Methods:

• Bronchoscope tracking is achieved solely through video data analysis.
• Topological localization within a generic airway model leverages sequential anatomical constraints.
• A hidden Markov model integrates CNN-based airway likelihoods from sequential frames.

Main Results:

• The vision-based approach demonstrated improved accuracy in a lung phantom model.
• Temporal context and anatomical knowledge regularization enhanced performance.
• Weighted F1 score improved significantly compared to frame-by-frame classification.

Conclusions:

• This research pioneers the combination of CNN image classification with HMM-based temporal inference for bronchoscopy guidance.
• The developed method shows promising results for vision-based guidance, independent of electromagnetic tracking and CT scans.
• This approach offers a potential solution for navigation in diverse clinical settings, including ICUs.