Description
Medical imaging research usually means labelling and training a new segmentation model for every organ or lesion. Medical-SAM3 adapts SAM3 into a medical foundation model that segments targets from a box or a short text prompt across many medical datasets.
The repo includes 2D benchmark inference and evaluation tools plus 3D training and inference code, with pretrained weights on Hugging Face for researchers to reproduce and build on.
Universal prompts:Box and text prompts with no per-task retraining.
2D and 3D:2D benchmark tooling and text-prompted 3D volume inference.
Open weights:Pretrained weights ready to download.
The repo includes 2D benchmark inference and evaluation tools plus 3D training and inference code, with pretrained weights on Hugging Face for researchers to reproduce and build on.
Features
Universal prompts:Box and text prompts with no per-task retraining.
2D and 3D:2D benchmark tooling and text-prompted 3D volume inference.
Open weights:Pretrained weights ready to download.
