Skip to content

Roadmap

This page tracks public-facing HemiSpec development after the v0.1.0 first public beta.

Current status

HemiSpec v0.1.0 was published on 2026-06-28 as a GitHub prerelease: https://github.com/mqqq333/HemiSpec/releases/tag/v0.1.0. It is research software / public beta, not a mature clinical or commercial product.

The archived release includes the hemispec CLI, hemispec-gui entry point, wheel/sdist, and Windows CLI/GUI artifacts. Current main adds the synthetic quickstart and model-cache download code. The public documentation targets that source checkout; these additions are not retroactive features of the archived release.

v0.1.x priorities

  1. Publish current functionality under a new version and immutable tag. Verify the resulting wheel, CLI, GUI, and synthetic quickstart independently, and preserve the archived v0.1.0 verification record.
  2. Validate inputs before inference: enforce the released-model grid and affine, reject duplicate subject/session keys, and check classifier atlas/features and optional dependencies before expensive work.
  3. Isolate outputs by run and make reruns safe: prevent stale subjects from entering summaries and restrict quickstart cleanup to its own generated files.
  4. Repair classifier-cache URLs, verify cached model hashes before loading, and publish versioned compatibility/provenance manifests for reusable assets.
  5. Add regression coverage for tied-value Spearman correlations, unsymmetrized TRT matrices, and ROI masks that retain valid zero residuals.
  6. Propagate GUI cancellation through classifier/TRT stages and improve missing-asset diagnostics.
  7. Add an optional study policy requiring hemisphere-classifier validation. Classification remains opt-in until that policy is implemented and tested; TRT still requires repeated scans.

v0.2 candidates

  • Publish hemispec-toolkit to PyPI after release metadata, ownership, and upload checks are complete; until then, continue GitHub Release distribution.
  • Zenodo DOI or equivalent archived software citation.
  • Richer atlas/custom-bundle downloader or resolver beyond the default released model cache.
  • Small approved demo dataset, if redistribution is permitted.
  • One-click HTML/PDF report generation.
  • Release CI that builds and uploads Windows artifacts automatically.
  • Stronger classifier/TRT output interpretation docs.
  • Manuscript, citation, and data-availability pages once the paper/archive DOI is public.

Non-goals for tracked source

Do not add unapproved model weights, atlas NIfTI files, real MRI inputs, generated outputs, private paths, or unpublished manuscript-only payloads to the source repository. The approved reusable DGN/classifier bundles under assets/models/ are the explicit Git-LFS exception; publish any additional assets separately with manifests and checksums.