name: mhub-segmentation description: Discover MHub medical imaging AI models, look up anatomical segment codes (SegDB/SNOMED), and generate workflow configurations for running models on NIfTI files. Use this skill when users ask about medical image segmentation models, what models can segment specific anatomy, how to run MHub models on their data, or need DICOM-SEG metadata. Works offline with cached data.

MHub Segmentation Skill

This skill provides tools for working with MHub medical imaging AI models and the SegDB anatomical segment database.

Capabilities

1. Model Discovery - Find models by modality, anatomy, or capability
2. Segment Lookup - Get SNOMED codes and metadata for anatomical structures
3. Workflow Generation - Create custom configs for NIfTI/NRRD input
4. DCMQI Config Generation - Generate metadata for DICOM-SEG conversion

Quick Reference

Find Models

# List all models
python scripts/mhub_helper.py models

# Filter by modality
python scripts/mhub_helper.py models --modality CT
python scripts/mhub_helper.py models --modality MR

# Find models that segment specific anatomy
python scripts/mhub_helper.py find liver kidney
python scripts/mhub_helper.py find heart cardiac
python scripts/mhub_helper.py find lung

# Get model details
python scripts/mhub_helper.py model totalsegmentator
python scripts/mhub_helper.py model lungmask

Look Up Segments

# Search segments
python scripts/mhub_helper.py segments --search kidney
python scripts/mhub_helper.py segments --search heart

# Get segment details (SNOMED code, color)
python scripts/mhub_helper.py segment LIVER
python scripts/mhub_helper.py segment LEFT_KIDNEY

Generate Workflow Configs

# Generate NIfTI workflow for a model
python scripts/mhub_helper.py config totalsegmentator --pattern flat --output custom.yml
python scripts/mhub_helper.py config lungmask --pattern subject_folders --output custom.yml
python scripts/mhub_helper.py config platipy --pattern bids --modality ct --output custom.yml

# Generate DCMQI config for DICOM-SEG
python scripts/mhub_helper.py dcmqi totalsegmentator --output dcmqi_meta.json

Scaffold Model Packages

python scripts/mhub_helper.py scaffold example_model \
   --label "Example Model" \
   --description "Basic segmentation package" \
   --modalities CT,MR \
   --primary-modality CT

Creates models/example_model with meta.json, config/default.yml, Dockerfile, and README.md copied from assets/model-templates so you can start customizing the package immediately.

Build Model Containers

python scripts/mhub_helper.py build example_model \
   --tag mhubai/example_model:latest \
   --build-arg VERSION=local_build

Runs docker build in models/example_model/ using the scaffolded Dockerfile, tagging it mhubai/example_model:latest by default and letting you supply --build-arg or --platform overrides.

Repository Root Overrides

If the skill lives outside the workspace (e.g., in .github/skills, .claude/skills, or a global install), point it at the real repo root with the top-level --repo-root /path/to/repo flag or the MHUB_SEGMENTATION_REPO_ROOT environment variable. This ensures scaffold and build treat models/ and the rest of the repo consistently no matter where the helper script is executed. The flag takes precedence over the environment variable; omit both to use the default heuristic.

Run Models via Docker

# Run a model on your NIfTI folder with an optional workflow or config override
python scripts/mhub_helper.py run lungmask \  
   --input /path/to/nifti \  
   --output /path/to/results \  
   --config ./custom.yml

If you prefer to use one of the built-in workflows, swap --config for --workflow default (or the workflow name from assets/workflow-templates).

Data Cache

This skill includes cached data for offline operation:

| Cache | Contents | Location | |-------|----------|----------| | Models | 30 MHub models with metadata | data/models_summary.json | | SegDB | 155 anatomical segments with SNOMED | data/segdb_cache.json | | Configs | Default workflows for all models | assets/workflow-templates/defaults/ |

Cache date: 2025-01-29

To refresh cache (requires network):

python scripts/mhub_helper.py refresh

Common Tasks

"What models can segment the liver?"

python scripts/mhub_helper.py find liver

Returns: totalsegmentator, nnunet_liver, bamf_nnunet_ct_liver, mrsegmentator, etc.

"How do I run TotalSegmentator on my NIfTI files?"

1. Generate a custom workflow config:

   python scripts/mhub_helper.py config totalsegmentator --pattern flat --output custom.yml
   

2. Run with Docker:

   docker run --rm --gpus all \
     -v /path/to/nifti:/app/data/input_data:ro \
     -v /path/to/output:/app/data/output_data \
     -v ./custom.yml:/app/config/custom.yml:ro \
     mhubai/totalsegmentator:latest \
     --config /app/config/custom.yml
   

For detailed NIfTI workflow instructions, see references/nifti-workflows.md.

"What's the SNOMED code for left kidney?"

python scripts/mhub_helper.py segment LEFT_KIDNEY

Returns: SNOMED code 64033007, color RGB(212, 126, 151)

"I need to create a DICOM-SEG from my segmentations"

1. Generate DCMQI metadata:

   python scripts/mhub_helper.py dcmqi totalsegmentator --output meta.json
   

2. Run DCMQI converter:

   itkimage2segimage \
     --inputImageList segmentation.nii.gz \
     --inputDICOMDirectory /path/to/dicom \
     --outputDICOM output.seg.dcm \
     --inputMetadata meta.json
   

File Organization Patterns

The skill supports three common file organization patterns:

| Pattern | Description | Example Structure | |---------|-------------|-------------------| | flat | All NIfTI in one folder | input_data/*.nii.gz | | subject_folders | One folder per subject | input_data/subject_id/*.nii.gz | | bids | BIDS-compliant | input_data/sub-XX/anat/*_T1w.nii.gz |

For advanced patterns (clinical trials, multi-site, custom naming), see:

• references/filestructure-patterns.md - FileStructureImporter syntax
• references/dataorganizer-patterns.md - Output organization options
• references/nifti-workflows.md - Complete workflow guide

Workflow Templates

Pre-built templates in assets/workflow-templates/:

| Template | Use Case | |----------|----------| | nifti_generic.yml | Starting point for any model | | bids_template.yml | BIDS-compliant data | | clinical_trial_template.yml | Multi-site with encoded filenames | | defaults/<model>.yml | Original DICOM configs for reference |

Available Models (30)

CT Segmentation

• totalsegmentator - 104 structures (organs, bones, muscles, vessels)
• platipy - 17 cardiac structures for radiotherapy
• lungmask - Lungs and 5 lobes
• casust - 8 cardiac structures
• nnunet_liver - Liver + tumor
• nnunet_pancreas - Pancreas + tumor
• bamf_nnunet_ct_kidney - Kidney + tumor + cyst

MR Segmentation

• mrsegmentator - 38 structures (CT/MR compatible)
• bamf_nnunet_mr_prostate - Prostate
• monai_prostate158 - Prostate zones
• gc_spider_baseline - 48 spine structures

PET/CT

• bamf_pet_ct_lung_tumor - Lung + FDG-avid tumor
• bamf_pet_ct_breast_tumor - Breast FDG-avid tumor

Prediction Models

• gc_picai_baseline - Prostate cancer likelihood
• gc_grt123_lung_cancer - Lung cancer risk
• pyradiomics - Radiomic feature extraction

Use python scripts/mhub_helper.py models for the complete list.

Dependencies

For offline use: No dependencies (uses cached data)

For cache refresh: pip install requests

For SegDB Python API: pip install segdb

Environment Notes

• Claude.ai (restricted network): Full functionality using cached data
• Claude Code: Full functionality + cache refresh capability
• Local execution: Full functionality + Docker for running models