MonoGID: geometry and illumination aware enhancement with distillation for self-supervised monocular endoscopic depth...

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Technology

Background:

• Monocular depth estimation is crucial for endoscopic procedures like 3D reconstruction and navigation.
• Endoscopic environments present challenges such as weak textures and variable lighting, hindering self-supervised learning methods.
• Existing methods often struggle with insufficient structural representation and limited prediction accuracy in these conditions.

Purpose of the Study:

• To improve self-supervised monocular depth estimation in endoscopic scenes.
• To address limitations of current methods in weak-texture and adverse illumination conditions.
• To enhance feature representation, prediction refinement, and deployment efficiency.

Main Methods:

• A task-specific architectural extension of the EndoDAC framework was developed.
• The method integrates geometry- and illumination-aware feature enhancement.
• Offline multi-generation self-distillation and inference-stage structural fusion were employed.

Main Results:

• The proposed EndoDAC extension improved depth estimation performance on the SCARED dataset.
• Lower error metrics were achieved on the Hamlyn dataset during zero-shot evaluation.
• Threshold accuracy remained comparable to the baseline on the Hamlyn dataset.

Conclusions:

• Task-specific adaptation of feature enhancement, self-distillation, and fusion improves endoscopic depth estimation.
• The architectural adaptation demonstrates effectiveness for self-supervised pipelines.
• Further research is needed to enhance cross-domain scale consistency and robustness in complex surgical scenarios.