A cGAN-based network for depth estimation from bronchoscopic images.

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Navigation

Background:

• Accurate 3D reconstruction of airway structures from 2D bronchoscopic images is crucial for developing vision-based navigation systems.
• Depth estimation is a foundational step for this 3D reconstruction.

Purpose of the Study:

• To enhance depth estimation performance directly from bronchoscopic images.
• To train a depth estimation network using both synthetic and real datasets.

Main Methods:

• A conditional Generative Adversarial Network (cGAN)-based network, BronchoDep-GAN, was developed.
• The network translates bronchoscopic images into depth maps.
• Training involved supervised learning on synthetic and virtual data, and unsupervised learning on unpaired real data.

Main Results:

• BronchoDep-GAN demonstrated superior accuracy in depth estimation compared to the established pix2pix cGAN.
• Performance improvements were observed on both synthetic and real bronchoscopic datasets.
• Real-data testing was qualitative due to the absence of ground truth.

Conclusions:

• Integrating virtual and real bronchoscopic images during training significantly improves depth estimation.
• Future work includes validation on 3D clinical phantoms and evaluating bronchoscope localization accuracy.
• The findings support advancements in vision-based bronchoscopic navigation systems.