Beyond Foundation Models: Distilling Geometric Priors for Lightweight Monocular Depth Estimation in Endoscopy

Area of Science:

• Computer Vision
• Medical Imaging
• Machine Learning

Background:

• Geometric foundation models excel in depth estimation but are computationally intensive.
• Surgical applications demand efficient, high-performance depth estimation models.
• Lightweight monocular depth estimators are crucial for real-time surgical guidance.

Purpose of the Study:

• To design a high-performance, lightweight monocular depth estimator for surgical applications.
• To transfer geometric knowledge from large foundation models into a compact network.
• To improve prediction accuracy and reduce artifacts in lightweight depth estimation.

Main Methods:

• Utilized geometric foundation models for their rich geometric priors.
• Introduced a novel trinity distillation scheme (spatial, spectral, gradient) for knowledge transfer.
• Developed a semantic distribution alignment strategy to mitigate pseudo-texture artifacts.

Main Results:

• The proposed lightweight estimator achieved state-of-the-art or comparable performance on SCARED, SERV-CT, Hamlyn, and C3VD datasets.
• Demonstrated a significantly smaller model size and reduced computational overhead compared to existing methods.
• Successfully suppressed pseudo-texture artifacts, enhancing prediction quality.

Conclusions:

• The novel trinity distillation and semantic alignment strategies enable efficient and accurate monocular depth estimation.
• The developed lightweight model is suitable for resource-constrained surgical applications.
• This research offers a promising solution for real-time geometric perception in surgery.