---

name: medical-sam-adapter description: "Use Medical-SAM-Adapter to prepare medical segmentation inputs, train SAM-family adapters, evaluate compatible checkpoints, and run the documented standalone MobileSAMv2 inference preflight." metadata: disco-role: operating disable-model-invocation: true license: GPL 3.0

Medical-SAM-Adapter

Route a medical image segmentation task through this skill when it depends on the Medical-SAM-Adapter command-line workflows, registered dataset adapters, SAM / EfficientSAM / MobileSAM model families, adapter or LoRA-style tuning, the repository checkpoint wrapper, or the separate MobileSAMv2 object-aware image path.

Route by the user's goal

• Inputs, dataset layout, checkpoints, tensor shapes, prompts, or 2D/3D preflight: read data-preparation.
• A new adaptation run or model/encoder/mode choice: read training, usually after data preparation.
• Periodic validation, independent val.py, checkpoint compatibility, metrics, visualization, or 3D evaluation chunking: read evaluation.
• Standalone object-aware MobileSAMv2 inference from images: read mobile-inference. Do not combine its decoder/detector weights with an adapter-training checkpoint without checking the separate contracts.

For a task spanning routes, use this order:

1. establish the user-selected environment, CUDA device, explicit data and checkpoint paths;
2. validate a representative sample and the exact case-sensitive dataset name;
3. select a complete model/encoder and adaptation mode, then construct a command without implicit downloads;
4. run a short, observable CUDA smoke before a long job;
5. inspect the produced checkpoint before independent evaluation; and
6. preserve logs and rendered outputs in explicit, separate directories.

Non-negotiable operating constraints

• The core training and evaluation code constructs CUDA tensors and moves models or data to CUDA. Standalone MobileSAMv2 also creates CUDA box tensors later in its path. A CPU import or metadata check is diagnostic only; CPU is not a substitute for the selected execution workflows.
• Do not install the legacy dependency file blindly: its CPU-only and CUDA entries conflict. Use a user-selected isolated environment and verify the actual dependency/backend combination before running.
• Checkpoint and dataset paths are user inputs. Do not download, infer, or overwrite them from a filename. Keep input data, checkpoints, logs, and rendered outputs separate.
• The source parser uses several type=bool arguments. A string such as False can become truthy in Python; inspect effective values rather than trusting command spelling.
• The bundled scripts are read-only preflight/inspection helpers. They do not launch training, evaluate real data, download artifacts, or perform actual MobileSAMv2 inference.

Public environment and bounded verification

The source repository has no standard package metadata. In a user-selected isolated Python environment, install a compatible CUDA-enabled PyTorch and the workflow dependencies required by the chosen route (for example MONAI, scikit-image, nibabel/SimpleITK, OpenCV, tensorboardX, and the matching MobileSAMv2 detector dependency when that optional route is selected). Do not blindly install the historical environment.yml: its CPU-only and CUDA pins conflict. Keep all dataset and checkpoint acquisition explicit and user-owned.

Before a long job, run the bundled, source-independent checks:

python scripts/inspect_cli.py --list
python sub-skills/training/scripts/inspect_model_registry.py --list
python sub-skills/data-preparation/scripts/validate_sample_contract.py --help
python sub-skills/evaluation/scripts/inspect_checkpoint.py --help
python sub-skills/mobile-inference/scripts/run_mobile_samv2.py --help

For actual training, evaluation, and MobileSAMv2 rendering, separately verify torch.cuda.is_available(), the selected device, and a small CUDA allocation. The commands above are safe contract checks, not proof of model execution.

Shared references and verification boundary

Read shared troubleshooting when an issue crosses environment, data, model, checkpoint, or output boundaries. The bundled helpers' --help paths and deterministic invalid-input behavior are safe checks. Full training, real-data evaluation, notebook execution, checkpoint/dataset downloads, and actual MobileSAMv2 rendering require external artifacts and are not implied by this skill's verification.

The source snapshot and evidence routes are recorded in repo provenance; routing metadata is in repo-routing-metadata.json.