Reproduce “Surgical-DINO: Adapter Learning of Foundation Models for Depth Estimation in Endoscopic Surgery” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Beilei Cui, Mobarakol Islam, Long Bai, Hongliang Ren

Surgical-DINO efficiently adapts a vision foundation model, DINOv2, for monocular depth estimation in endoscopic surgery by using Low-Rank Adaptation (LoRA). The method achieves state-of-the-art performance on the SCARED dataset and demonstrates robust zero-shot generalization on the challenging Hamlyn dataset, requiring only 0.17% of the foundation model's parameters to be trainable.

Key insights

  • Generalist foundation models applied zero-shot to surgical scenes perform poorly, confirming a substantial domain gap that necessitates targeted adaptation.
  • Efficient adaptation strategies like LoRA are critical for successfully deploying large foundation models in resource-constrained, specialized medical domains.
  • Adapting the encoder's core visual feature representations, rather than just training a separate decoder, is crucial for achieving high accuracy in surgical depth estimation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.