| name | ms-lesion-segmentation |
| description | Workflow for multiple-sclerosis (MS) white-matter lesion segmentation and lesion-load quantification from a T1w + FLAIR pair using LST-AI |
| license | Apache-2.0 |
| compatibility | Requires the medmcp-neuro-ms MCP server (console script medmcp-neuro-ms). |
MS lesion segmentation & lesion-load workflow
segment_ms_lesions runs LST-AI — a deep-learning UNet3D ensemble (native PyTorch) — to
segment MS white-matter lesions and report total lesion load (volume mm³), lesion
count, and a per-region breakdown. It does its own registration to MNI
(picsl-greedy) and skull stripping (HD-BET), so you only provide the raw scans.
When to use
- The user wants MS lesion segmentation, a lesion mask, or lesion load /
lesion volume / lesion count from a brain MRI.
- The user has both a T1w and a FLAIR of the same session.
Required inputs
LST-AI needs both a T1w and a FLAIR, co-registered to the same session. If only
one contrast is available, tell the user LST-AI cannot run (it is a T1+FLAIR method) —
do not substitute another modality.
Steps
- Identify the T1w and the FLAIR. Match by contrast, not by guessing — the filename
usually contains
T1w / FLAIR. If you cannot tell which is which, ask.
- Do not skull-strip or register first. LST-AI handles both internally. Only pass
already_stripped=True if the user states both images are already skull-stripped.
- Run
segment_ms_lesions with device="auto" (GPU when present, else CPU — tell
the user if it resolves to CPU, which is slower). Keep annotate=True (default) unless
the user only wants the binary mask.
- Report lesion load. From the result:
total_lesion_volume_mm3, lesion_count,
and the lesion mask path. If region_volumes_mm3 is present, give the per-region
breakdown. Follow the result's _render rules.
Region annotation
With annotate=True, lesions are labelled into four McDonald-criteria regions —
Periventricular, Juxtacortical, Subcortical, Infratentorial — written to the
annotated map and *_annotated_lesion_stats.csv. Call list_ms_lesion_regions() for
the exact region names.
Gotchas
- Research only. LST-AI is not validated/approved for clinical use. If the user
describes a clinical decision context, flag this clearly before proceeding.
- Output is a binary mask in FLAIR space (
*_seg-lst.nii.gz) — offer to overlay it
on the FLAIR in the viewer. The annotated map (*_desc-annotated_seg-lst.nii.gz) uses
labels 1–4 for the four regions.
- Lesion load comparisons over time need consistent acquisition; raw mm³ differences
between scans can reflect protocol/scanner changes, not true disease change — caveat any
longitudinal comparison.
- Runtime — minutes on GPU; substantially longer on CPU. The result reports which
device was used.
- Errors: report and stop; do not retry with modified inputs without asking the user.
