A Disparity Refinement Framework for Learning-based Stereo Matching Methods in Cross-domain Setting for Laparoscopic...

Area of Science:

• Computer Vision
• Medical Imaging
• Surgical Technology

Background:

• Depth estimation via stereo matching is vital for computer-assisted surgery (CAS) visualization.
• Learning-based stereo matching offers promise but suffers from data requirements and domain shift issues.
• Robustness and performance improvements for learning-based methods remain challenges.

Purpose of the Study:

• To propose a disparity refinement framework to enhance learning-based stereo matching for cross-domain applications.
• To improve the accuracy and robustness of depth estimation in CAS, particularly with laparoscopic imagery.
• To address limitations of current methods when applied to unseen datasets.

Main Methods:

• Developed a two-part disparity refinement framework: local and global methods.
• Applied the framework to refine results from pre-trained learning-based stereo matching models.
• Tested the framework on laparoscopic image datasets, evaluating performance on unseen data.

Main Results:

• The proposed framework effectively refines disparity maps on unseen datasets, even with noise.
• Performance is maintained without compromising correct predictions when networks generalize well.
• Demonstrated potential for robust and accurate disparity prediction in cross-domain settings.

Conclusions:

• The disparity refinement framework enhances learning-based stereo matching for CAS.
• It offers a solution for improving depth estimation accuracy and robustness, especially in the absence of large, specific training datasets.
• Incorporating this framework into existing networks is beneficial for advanced surgical visualization.