BronchoGAN: anatomically consistent and domain-agnostic image-to-image translation for video bronchoscopy

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Limited availability of bronchoscopy images hinders deep learning model training.
• Robust image translation across virtual, phantom, in vivo, and ex vivo domains is crucial for clinical applications.

Purpose of the Study:

• To develop BronchoGAN, a novel approach for synthesizing realistic bronchoscopy images.
• To enable robust image translation across diverse bronchoscopy data domains.
• To generate large-scale, realistic bronchoscopy image datasets for training deep learning models.

Main Methods:

• Integration of anatomical constraints, specifically bronchial orifice matching, into a conditional Generative Adversarial Network (GAN).
• Utilizing foundation model-generated depth images as an intermediate representation for enhanced robustness and paired data construction.
• Employing image-to-image translation techniques for domain adaptation.

Main Results:

• Successful translation of input images from various domains (virtual bronchoscopy, phantoms) to realistic human airway appearance.
• Robust preservation of anatomical features, such as bronchial orifices, demonstrated qualitatively and quantitatively.
• Significant improvements in image quality metrics including FID, SSIM, and Dice coefficients (up to 0.43 improvement for synthetic images).

Conclusions:

• BronchoGAN effectively translates images across different bronchoscopy domains using intermediate depth representations and anatomical constraints.
• The approach enables the incorporation of public CT scan data (virtual bronchoscopy) to create extensive, realistic datasets.
• BronchoGAN addresses the scarcity of public bronchoscopy images, facilitating advancements in AI-driven medical diagnostics.