Inspected the exact histo-v0.0.127 source tag (b31b1223) in the local Histo checkout.
Documented the public box-level v2 tissue head loader and runner, including FeatureGrid, RGB, normalization, and output requirements.
Compared that path with SemanticSegPredictor whole-slide APIs and the full domain_representation.process_scan orchestration.
Recommended that NanoVirDx delegate to process_scan for v2 domain representations rather than rebuilding embedding, windowing, merging, or persistence.
Confirmed that standalone semantic whole-slide prediction returns arrays and has no production persistence helper; full domain persistence belongs to Histo’s internal WsiOutputs path.
No source files in the target monorepo were edited.
Lessons Learned: Pitfalls
The local Histo checkout did not initially have the histo-v0.0.127 tag. Searching commits by package version found a merge from another branch, not the released commit. Querying remote tags and fetching the exact tag was necessary.
The term “tissue segmentation” can refer to the v2 shared-embedding tissue head or the older/general semantic whole-slide predictor. They have different model and input contracts and must not be treated as interchangeable.
Lessons Learned: Improvements
The Histo SOP should name the released tag pattern (histo-vX.Y.Z) and distinguish box-level head APIs, standalone semantic WSI inference, and full domain-representation orchestration.
A concise compatibility table for loaders, input ownership, outputs, and persistence would make future NanoVirDx adapter work faster and reduce accidental orchestration duplication.
