Leveraging Spatiotemporal Cues for Self-Supervised Stereo Depth Estimation in Endoscopic Videos.

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Technology

Background:

• Current self-supervised endoscopic depth estimation methods often ignore temporal dynamics, leading to artifacts.
• This limitation hinders accurate depth reconstruction in dynamic surgical environments.

Purpose of the Study:

• To develop a robust self-supervised framework for accurate depth estimation in endoscopic videos.
• To address limitations of existing methods by incorporating temporal and stereo information.

Main Methods:

• Introduced a Siamese self-supervised framework utilizing stereo and temporal data.
• Developed Deformable Spatiotemporal Cross-View Fusion (DSCF) and Multi-Scale Selective Spatiotemporal Aggregation (MSSA) modules.
• Implemented a three-stage training protocol for efficient edge deployment and monocular depth estimation.

Main Results:

• Achieved state-of-the-art performance on four public endoscopic datasets (SCARED, SERV-CT, EndoNeRF, Hamlyn).
• Demonstrated significant improvements in depth accuracy and image reconstruction quality.
• Reduced root mean square error (RMSE) by over 9.1% compared to existing methods.

Conclusions:

• The proposed framework effectively enhances endoscopic depth estimation by integrating spatiotemporal information.
• The method offers a promising solution for real-time surgical applications and edge device deployment.