Confidence-Guided Fusion for Self-Supervised Monocular Depth Estimation in Endoscopy

Area of Science:

• Computer Vision
• Medical Imaging
• Surgical Technology

Background:

• Accurate monocular depth estimation (MDE) is crucial for endoscopic surgery, improving navigation and depth perception.
• Existing diffusion and discriminative models have complementary strengths but also asymmetric error patterns.
• Discriminative models excel at geometric boundaries but fail in homogeneous areas; diffusion models capture texture but may lack structural coherence.

Purpose of the Study:

• To develop a novel framework, CoDepth, that harmonizes heterogeneous depth estimators by leveraging confidence-guided fusion.
• To systematically exploit the complementary strengths of diffusion-based and discriminative depth estimation models.
• To improve the accuracy, robustness, and generalizability of MDE in endoscopic surgery.

Main Methods:

• Introduced CoDepth, a framework employing confidence-guided fusion to integrate outputs from diverse depth estimators.
• Developed a complementary map extractor to identify disparity disagreements between models.
• Integrated a cross-attention module for context-aware feature integration and a probabilistic confidence network for adaptive fusion weights.

Main Results:

• CoDepth demonstrated improved overall performance on the SCARED dataset compared to single-model baselines, particularly in Abs Rel and δ-based accuracy.
• The framework showed encouraging cross-domain generalization, achieving competitive performance on SERV-CT, Hamlyn, and C3VD datasets without fine-tuning.
• CoDepth exhibited enhanced robustness against synthetic corruptions like low-light, Gaussian noise, and impulse noise.

Conclusions:

• Confidence-guided complementary fusion offers a practical integration paradigm for combining heterogeneous endoscopic depth estimators.
• CoDepth enhances MDE accuracy and robustness, showing significant potential for improving surgical navigation and clinical utility.
• The framework's cross-domain generalization and robustness highlight its adaptability to complex and varied surgical environments.