| name | medsam2-3d-segmentation |
| description | Operate MedSAM2 for promptable segmentation of 3D medical images and medical videos, including CT lesion propagation, MRI volumes, RECIST-guided prompts, efficient CPU-oriented variants, training, and 3D Slicer integration. Use when generating or validating volumetric masks from sparse prompts or propagating masks through image slices or video frames. |
MedSAM2 3D Segmentation
Use MedSAM2 to accelerate annotation and research segmentation. Every generated mask requires anatomical and task-specific quality review before quantitative or clinical use.
Workflow
- Define the target anatomy or lesion, modality, acquisition protocol, prompt source, and downstream measurement.
- De-identify inputs and preserve voxel spacing, orientation, affine, frame order, and original series identifiers in a controlled manifest.
- Convert DICOM or vendor formats with a validated pipeline. Do not treat array index order as physical orientation.
- Choose the prompt strategy:
- point or box prompt for interactive segmentation;
- RECIST marker or middle-slice box for lesion propagation;
- initial mask for video propagation.
- Run a pilot and inspect propagation across the full volume or video, not only the prompted slice or frame.
- Post-process only with prespecified operations such as component filtering or hole filling, and compare masks before and after each operation.
- Evaluate Dice, surface Dice, Hausdorff distance, lesion-wise detection, volume error, and inter-reader variability as appropriate.
- Stratify performance by scanner, site, contrast phase, slice thickness, anatomy, lesion size, and prompt quality.
- Export masks with the original geometry and retain prompt provenance for audit.
Guardrails
- Do not use unreviewed masks for diagnosis, radiation planning, surgery, or response assessment.
- Detect missing slices, inconsistent orientation, anisotropic spacing, motion, and corrupted frames.
- Review small lesions and volume boundaries slice by slice.
- Do not evaluate only on datasets used for model development.
- Verify checkpoint, dataset, SAM2, and downstream plugin licenses separately.
- Require qualified clinical annotation and local validation for real-world deployment.
Output Contract
Return input geometry and provenance, prompt type, checkpoint, preprocessing, segmentation metrics, per-case failures, subgroup analysis, post-processing, exported-mask format, and review status.
Read references/operations.md for installation, inference commands, training patterns, and canonical sources.