Monocular absolute depth estimation from endoscopy via domain-invariant feature learning and latent consistency

Area of Science:

• Medical Robotics
• Computer Vision
• Surgical Navigation

Background:

• Monocular depth estimation (MDE) is crucial for guiding autonomous medical robots in surgery.
• Acquiring precise metric depth from endoscopic cameras in surgical settings is challenging, hindering supervised learning on real data.
• Existing unsupervised domain adaptation methods for MDE often leave a domain gap between real and translated synthetic images.

Purpose of the Study:

• To develop a latent feature alignment method to improve absolute depth estimation in endoscopic videos.
• To reduce the domain gap between synthetic and real endoscopic images for more accurate depth prediction.
• To enhance the performance of depth estimation networks by focusing on latent feature learning.

Main Methods:

• A latent feature alignment method is proposed, agnostic to image translation processes.
• The depth network learns domain-invariant features using adversarial learning and directional feature consistency.
• The method processes both translated synthetic and real endoscopic frames to refine depth estimation.

Main Results:

• The proposed method demonstrates superior performance in absolute and relative depth metrics compared to state-of-the-art MDE techniques.
• Consistent improvements were observed across various network backbones and pre-trained weights.
• The approach effectively reduces the domain gap in endoscopic depth estimation.

Conclusions:

• Latent feature alignment offers a robust solution for improving absolute depth estimation in challenging endoscopic scenarios.
• The method enhances the reliability and accuracy of depth information for autonomous surgical systems.
• This work contributes to advancing MDE for safer and more precise robotic surgery.