Reproduce “Medical Image Segmentation: Foundation Model Adaptation versus Task-Specific Training” 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

Emre Kutlusan, Yongmin Li

The authors compare adapting a pretrained DINOv2-Small encoder with training U-Net-style segmentation models from scratch across polyp, breast ultrasound, and skin lesion datasets. They use LoRA and evaluate held-out source images and external test sets. Adaptation has higher DSC on all eight external sets, but two differences fall within measurement noise; task-specific models lead on fine-boundary scores on four sets.

Summary by alphaXiv.

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