Inspected histo-v0.0.127 and origin/main in the private Histo repository. The packaged Histo source is identical between those refs.
Identified histo.semantic_segmentation.inference.SemanticSegPredictor as the public whole-slide API and documented its checkpoint loading, device behavior, predict_wsi signature, normalization inputs, and in-memory uint8 outputs.
Confirmed the API is included in the 0.0.127 wheel and that the conda recipe carries its GPU/runtime dependencies.
Proposed a thin NanoVirDx adapter that owns path opening and output persistence only, while delegating model inference and sliding-window orchestration to Histo.
Identified missing policy decisions around target MPP, normalization, tissue-mask requirements, and ClearML model ID or local checkpoint configuration.
Lessons Learned: Pitfalls
The Histo Pixi CUDA feature still lists nnunetv2, but current packaged source does not import it. The authoritative runtime dependency check for consumers is the conda recipe, not the development Pixi feature list.
A historical WSI runner contains useful reference orchestration but was removed and was never packaged. It must not be treated as a supported import.
Lessons Learned: Improvements
Add a Histo SOP for the public standalone Gleason WSI contract, including the expected 1.0 µm/px input, normalization recipe, model ID source, label IDs, and an OME-Zarr path-to-path example.
Document whether NanoVirDx receives preprocessed single-resolution components or arbitrary WSI pyramids. This determines whether level selection belongs in the adapter.
