CUHK-AIM-Group/UN-SAM

UN-SAM addresses a critical bottleneck in using Meta's SAM for medical imaging: the requirement for manual prompts (points/boxes). By introducing a self-prompting mechanism and domain adaptation, it makes SAM viable for high-throughput pathology. However, its defensibility is limited (score 4) because it is primarily a research implementation on top of a rapidly evolving foundation model. With the release of SAM 2 and subsequent iterations from frontier labs, the specific architectural modifications in UN-SAM risk being superseded by more robust, natively multi-modal or zero-shot foundation models. Quantitatively, 90 stars and 8 forks after two years indicate it is a respected academic contribution (published in MedIA'25) but lacks the 'infrastructure-grade' traction needed for a higher score. It competes with established specialized tools like HoVer-Net, StarDist, and Cellpose, which have deeper clinical integration and larger community footprints. The displacement risk is medium-high over a 1-2 year horizon as foundation models for biology (like those from Recursion or specialized medical AI startups) continue to mature.