MultiverSeg (multiverseg.csail.mit.edu)

MultiverSeg is an interactive segmentation system designed to let researchers and clinicians rapidly label entirely new biomedical imaging datasets without any pre-existing labels from that task or domain. Instead of training from scratch, the model takes as input the target image, optional user interactions (clicks, bounding boxes, scribbles), and a growing context set of previously segmented image–segmentation pairs. As a user labels more images, those examples become context that the model conditions on, so the number of interactions required per image falls—amortizing human effort across the dataset. Technically, MultiverSeg uses a UNet-style encoder–decoder and a CrossBlock mechanism with extra normalization layers to fuse features from the target inputs and the context set throughout the network. The target stack can also include a prior predicted mask to support iterative corrections. On unseen segmentation tasks, MultiverSeg reduced total clicks by 36% and scribble steps by 25% to reach 90% Dice compared to a state-of-the-art interactive baseline, demonstrating that the model effectively leverages a small context set (e.g., 10 examples) to improve predictions. The approach enables few-shot, context-aware interactive segmentation that can speed annotation, reduce labeling cost, and adapt quickly to new biomedical imaging domains.