New AI system could accelerate clinical research (news.mit.edu)

MIT researchers unveiled MultiverSeg, an interactive AI system that dramatically speeds up segmentation of biomedical images by combining user interactions (clicks, scribbles, boxes) with an internal “context set” of previously segmented images. Unlike task-specific models that require hundreds of labeled examples and retraining, or prior interactive tools that force users to repeat marking for every image, MultiverSeg incrementally learns from the examples it creates during use. As the user marks more images, the model needs fewer corrections—eventually zero—and can accurately segment new images without retraining or machine-learning expertise. The work is from MIT CSAIL and collaborators at Harvard/MGH and was presented at ICCV. Technically, MultiverSeg’s architecture is built to reference a context set of any size, letting it adapt quickly across modalities (for some X‑rays it needs only one or two manual segmentations). Compared with state-of-the-art interactive and in-context baselines, it outperforms them and reduces user burden: by the ninth image it required only two clicks to beat a task-specific model, and it hit 90% accuracy with roughly 2/3 the scribbles and 3/4 the clicks of the team’s previous ScribblePrompt system. Practical implications include faster clinical research, lower cost and time for trials, and improved clinical workflows like radiation planning; next steps are real‑world clinical testing and extending the method to 3D imaging.