ilex-hub/sam_med3d.turbo.1

SAM-Med3D (promptable 3D medical-image segmentation) -- SAM-Med3D-turbo (ViT-B/16 3D, point-promptable)

Description

SAM-Med3D, ported to JAX / Equinox from the upstream PyTorch release. SAM-Med3D is a fully-3D re-expression of Meta's Segment Anything: a ViT-B/16 3D image encoder (windowed + global attention with decomposed relative positions), a point-prompt encoder (random-Fourier positional encoding + learnable point embeddings), and a two-way-transformer mask decoder that cross-attends prompt tokens and image tokens to predict masks + an IoU-quality estimate. This port wraps the released sam_med3d_turbo checkpoint. Forward: one volume + a set of point prompts -> low-resolution mask logits (quarter resolution) and per-mask IoU predictions.

Intended use

Promptable 3D segmentation of a single-channel medical volume. Inputs are the volume (normalised + zero-padded to 128^3 inside the model) and a set of point prompts -- voxel coordinates (N, 3) plus labels (N,) where 1 = foreground, 0 = background. With multimask output it returns 3 candidate low-resolution mask-logit volumes (32^3 = quarter resolution) and 3 scalar IoU-quality predictions (1 of each otherwise). v0 exposes point prompts only; box / mask prompts and the full-resolution interpolate+threshold postprocess are deferred. Resample / crop to the prompt (the upstream SamPredictor pipeline) before calling.

Usage

from ilex.models.sam_med3d import SAMMed3D
model = SAMMed3D.from_pretrained('ilex-hub/sam_med3d.turbo.1')

Authors

Wang H., Guo S., Ye J., et al.

Citation

Wang H., Guo S., Ye J., Deng Z., Cheng J., Li T., et al. (2023). SAM-Med3D. arXiv:2310.15161. Built on Kirillov A., et al. (2023), Segment Anything, ICCV 2023, arXiv:2304.02643.

References

• Wang H., Guo S., Ye J., et al. (2023). SAM-Med3D. arXiv:2310.15161.
• Kirillov A., Mintun E., Ravi N., et al. (2023). Segment Anything. ICCV 2023. arXiv:2304.02643.
• Weights + code: https://github.com/uni-medical/SAM-Med3D

License

HF Hub license tag: apache-2.0

Effective terms: Apache-2.0 (the SAM-Med3D authors, uni-medical) on both the 3D Segment Anything network code and the released sam_med3d_turbo weights. The underlying Segment Anything design is Meta's (also Apache-2.0). No commercial restrictions; no gating required. The ilex JAX / Equinox port code is separately licensed under Apache-2.0 / GPL-3.0.

Upstream license reference: https://github.com/uni-medical/SAM-Med3D/blob/main/LICENSE

Copyright

Network architecture (3D Segment Anything) and pretrained weights: copyright (c) the SAM-Med3D authors, released under the Apache-2.0 License. JAX / Equinox port: copyright (c) the ilex authors, released under the Apache-2.0 / GPL-3.0 dual license used by ilex itself.

Upstream source

Original weights / reference implementation: https://github.com/uni-medical/SAM-Med3D

Provenance

This artefact was produced by ilex's save/load pipeline. The architecture is implemented in ilex.models.sam_med3d.SAMMed3D and the weights have been converted from their upstream format. See the upstream source above for the canonical reference.