Researchers compare adapting DINOv2, a foundation model pretrained on natural images, for medical image segmentation against task-specific training. They vary how much of the model is tuned and test performance at new clinical sites, rather than only on training-distribution data.
Why it matters
It tests whether foundation models genuinely generalise to new hospitals or merely perform well where they were trained.
Analysis
Medical AI may favour targeted adaptation and site-robust evaluation over assuming large pretrained models replace specialised training.
Source: arXiv
AI disclosure: this article was produced with AI assistance and reviewed for accuracy.
