Enhanced self-supervised monocular depth estimation with self-attention and joint depth-pose loss for laparoscopic...

Area of Science:

• Computer Vision
• Medical Imaging
• Surgical Navigation

Background:

• Depth estimation is crucial for laparoscopic surgery navigation.
• Existing self-supervised methods struggle with textureless organs and complex camera rotations.
• Accurate depth and pose estimation are challenging in laparoscopic environments.

Purpose of the Study:

• To propose a novel self-supervised monocular depth estimation method for laparoscopic images.
• To enhance depth and pose estimation accuracy in challenging surgical scenes.
• To introduce a self-attention-guided pose estimation and a joint depth-pose loss function.

Main Methods:

• Extracted feature maps and calculated minimum re-projection error for feature-metric loss.
• Incorporated self-attention blocks into the pose estimation network.
• Developed a joint depth-pose loss function combining feature-metric and pose losses.

Main Results:

• Achieved significant improvements in absolute relative error on SCARED and Hamlyn datasets (18.07% and 14.00%).
• Generated smooth depth maps with low error in diverse laparoscopic scenarios.
• Demonstrated a favorable trade-off between computational efficiency and performance.

Conclusions:

• The proposed method effectively addresses laparoscopic depth estimation challenges.
• All introduced components significantly contribute to the method's performance.
• The method offers an efficient balance between computational cost and accuracy for surgical navigation.