An Enhanced Synthetic Cystoscopic Environment for Use in Monocular Depth Estimation

Area of Science:

• Medical Imaging
• Computer Vision
• Robotic Surgery

Background:

• Accurate device positioning is crucial for minimally invasive surgeries.
• Current optical localization methods struggle in dynamic, non-rigid surgical environments like cystoscopies.
• Monocular endoscopic camera images present challenges for precise tracking.

Purpose of the Study:

• To develop an improved method for supervised depth estimation using neural networks.
• To reduce the domain gap between synthetic and real endoscopic images.
• To enhance the accuracy of device positioning during manual, minimally invasive procedures.

Main Methods:

• Utilized neural networks for supervised depth estimation from synthetic images.
• Employed adversarial training in a second step to apply the network to real images.
• Enhanced a synthetic cystoscopic environment to minimize the domain gap.

Main Results:

• The proposed enhanced synthetic environment demonstrated distinct improvements over previous work.
• The trained network showed superior performance when applied to real test images.
• Achieved more accurate device positioning in simulated and real endoscopic scenarios.

Conclusions:

• Improved synthetic data generation and adversarial training enhance depth estimation accuracy.
• The developed method offers a more robust solution for device tracking in minimally invasive surgery.
• This approach has the potential to improve surgical precision and patient outcomes.