BiomedParse

BiomedParse is Microsoft Research's biomedical image foundation model, published November 18, 2024 in Nature Methods with collaborators at the University of Washington and Providence. It unifies object recognition, detection, and segmentation in a single framework: instead of drawing bounding boxes, a clinician types a plain-English description like pathological cells and the model finds and segments the matching structures.

Version 1 covers 2D images across nine modalities (CT, MRI, ultrasound, X-ray, pathology, endoscopy, dermoscopy, fundus, and OCT) with 100+ tasks. Version 2, released in October 2025, added 3D volume support, and the model took first place in the CVPR 2025 challenge on foundation models for text-guided 3D biomedical image segmentation.

The honest caveats: it is a research model, not a clinical device, and Microsoft's own disclaimer stresses it must not substitute for professional medical judgment. Code and weights are published on GitHub; check the repository for the exact license terms.