write-imaging-factsheet-ehr

Write Imaging Factsheet for EHR

Use the four medimageparse MCP tools; VS Code may display them as mcp_medimageparse_<tool>.

Required tool sequence

Perform these calls for the supplied images in this request. Loading this skill or reading an existing audit folder is not a substitute.

1. Resolve image_path and any prior_image_path to absolute paths.
2. Call list_models once. Select model mip3d-mri-brain-lesion and its views["tumor-components"].prompts. Pass the full prompts verbatim, in their returned order.
3. Call segment for the current image with that model and those prompts. Keep its audit_dir, mask_path, and review-image paths. Call clean_segmentation with that returned mask_path, using the default cleanup parameters; keep the returned clean_path.
4. If a prior image was supplied, call segment and then clean_segmentation for it in the same way, with the same model, prompts, and cleanup parameters. Keep the prior paths separately; never swap them with the current paths.
5. Call describe_lesions once with the current clean_path. Also pass prior_clean_path when a prior was supplied; omit it otherwise. Keep descriptors_path, inputs, conventions, descriptors, and the returned review-image paths.
6. Fill template/factsheet.md from the returned current descriptors and aggregate change descriptors. Save it beside descriptors_path, replacing .descriptors.json with .factsheet.md. Preserve all masks, sidecars, and descriptor JSON files.

If any required tool is unavailable or fails, stop and identify the step, image, and error; tell the user to check the medimageparse server or the reported input problem. Do not return a success-shaped factsheet or allow the parent to emit <report>. Do not run the server's Python code, write scripts, compute your own measurements, or substitute an earlier run.

Filling and checking the factsheet

• Replace placeholders with actual descriptor IDs and values. Repeat the lesion, pair, volume, change, and active-limitation lines for all applicable descriptors. Cite an ID on every fact line.
• For every current lesion, copy both diameters_cm entries from L*.bidimensional and both value.AP and value.TR from L*.extent. The bidimensional value and unit describe the product in cm2, not a linear dimension. Never use unrounded_diameters_cm.
• Include the enhancing, necrotic, and tumor-core volumes for every current lesion, all returned total volumes, the lesion count, and sum of products. With a prior, include every change.* line with its prior/current values, units, percent change, and any reason.
• Include returned diameters even when measurable is false. That flag is not permission to omit dimensions.
• Copy numbers, units, categories, booleans, and reasons verbatim. For explicit unavailable values, preserve null and the reason. Never fill a missing field with zero or infer a reason.
• A successful zero-lesion result has no per-lesion descriptors: omit those sections but keep the zero count, totals, change lines when present, and limitations. It is not a tool failure.
• Do not compute, round, convert, diagnose, assign anatomy or side, or match lesion IDs across scans. Do not copy historical report sizes into this factsheet.
• Copy frontmatter from inputs and conventions, using valid YAML quoting for paths and literal null for a missing prior. Prior lesion descriptors are saved at inputs.prior_descriptors_path; the returned lesion inventory is the current scan's, not a matched current/prior inventory.
• Link the current scan's review images by filename in the Images section. Omit its measurements-image line if no measurement image was returned for the current scan, even if an image was returned for the prior.
• Before returning, check every current lesion has its diameter pair, AP/TR pair, and component volumes, and that required totals and applicable changes are present. Repair omissions from the same descriptor result or its saved JSON, not by rerunning segmentation. If a required source field is missing, stop with an explicit error naming it.

Handoff

• When called by write-report-with-ehr: retain the factsheet and pass its contents, saved path, descriptors_path, and each scan's audit_dir back to the parent workflow, current first. Continue to the parent's report-writing step. Do not send the factsheet as the final response or end the task with audit folders.
• When explicitly requested on its own: reply with the completed factsheet, then its saved path and each scan's audit-log folder, current first.