Disparity refinement framework for learning-based stereo matching methods in cross-domain setting for laparoscopic...

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Technology

Background:

• Depth estimation is vital for computer-assisted surgery visualization.
• Learning-based stereo matching shows promise but suffers from domain shift and data requirements.
• Robustness and accuracy of these methods remain challenges.

Purpose of the Study:

• To develop a disparity refinement framework to enhance learning-based stereo matching for laparoscopic surgery.
• To address limitations of existing methods, including domain shift and data dependency.
• To improve the accuracy and robustness of depth estimation in cross-domain settings.

Main Methods:

• A disparity refinement framework combining local and global refinement was proposed.
• Learning-based stereo matching methods pre-trained on natural images were tested on laparoscopic datasets.
• The framework was evaluated for its ability to refine noisy disparity maps and maintain accuracy.

Main Results:

• The proposed framework effectively refines noise-corrupted disparity maps on unseen datasets.
• Prediction accuracy was maintained when the network generalized well.
• Qualitative and quantitative results demonstrated the framework's efficacy.

Conclusions:

• The disparity refinement framework enhances robustness and accuracy of learning-based stereo matching for depth estimation.
• It offers a solution for improving performance despite limited laparoscopic training data.
• Integration into existing networks can significantly boost depth estimation capabilities in surgery.