Foundation Models for Medical Imaging
Building and fine-tuning vision foundation models for clinical imaging tasks.
We develop and benchmark foundation models tailored for medical imaging tasks. Our work spans:
- General-purpose segmentation models like SegmentAnyBone and SegmentAnyMuscle, adapted from SAM.
- MRI-specific encoders, such as MRI-CORE, self-supervised trained on 116,806 volumes.
- Zero-shot and cross-modality generalization, evaluating foundation models in zero-shot segmentation and registration tasks.
Our models aim to reduce annotation cost and improve generalizability in real-world clinical settings.
Left: Bone segmentation in knee MRI. Right: Multi-sequence MRI model using MRI-CORE.