Conference paper

RAPTOR: Scalable Train-Free Embeddings for 3D Medical Volumes Leveraging Pretrained 2D Foundation Models

PaperCode

*Equal contribution; Spotlight (top 2.6% of submissions)

Pipeline: a stack of scan slices passes through a frozen two-dimensional image model and a random projection into a short embedding vector.
RAPTOR: no training, one consumer graphics card, about 6.5 seconds per scan.

A train-free framework for 3D medical volumes that led the evaluated methods on 7 of 10 datasets, processed a full volume in about 6.5 seconds on a single 11 GB GPU, and enabled brain-MRI analyses identifying 126 new genetic associations while replicating 111 of 118 previously reported associations.

All publications