Reproduce “Impact of Pretraining Length, Model Size, and Annotation Budget on Downstream Clinical Tasks of Foundation Models in Radiotherapy” 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

Alejandro Cortina Uribe, Sidsel Højklint Poulsen, Ivan Richter Vogelius, Jens Petersen

The authors pretrain vision transformers on routine three-dimensional cone-beam CT (CBCT) scans by hiding image patches and reconstructing them, then adapt the encoders to labeled tasks with full fine-tuning or LoRA. For feeding-tube classification, fully fine-tuned pretrained models reach AUROC 0.91–0.97, versus 0.58–0.85 without pretraining; these are ranges across models in retrospective, single-center data, not evidence of patient benefit.

Summary by alphaXiv.

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