feat: deliver end-to-end research imaging workbench

[Liohtml]

Summary

The browser wizard previously ended at a 501 Not Implemented response. This PR connects it to a complete local research workflow: upload DICOM data, select a study, run an analysis in the background, inspect slices and findings, review results, and download reports.

It also fixes data-processing issues that could mix series, misorder slices, omit later series from vision analysis, or discard processing warnings. The implementation is a coordinated workbench upgrade spanning the pipeline, API, UI, provider integrations and deployment documentation.

Fixes #157. Related: #61, #25, #57, #58, #18, #54.

Changes

• Explicitly close viewer memory maps before deletion/shutdown, copy slices under the store lock, and keep failed deletions visible for retry.
• Add authenticated upload/inspection APIs, restricted server paths, bounded background jobs, progress, cooperative cancellation, persistent results and memory-mapped viewer images, report downloads and explicit data deletion.
• Improve the four-language browser workflow, mobile layout, keyboard accessibility, error recovery, reload/reconnect behavior, image-reference navigation and finding edits with before/after audit history.
• Group DICOM by study/series UIDs; support individual images and DICOMDIR; sort using geometry and expose data-quality warnings. Fix ResNet evaluation mode/batching, signal metric semantics, balanced vision image selection and preservation of warnings.
• Add conservative metadata de-identification, identifier cleanup, optional local OCR/manual masks, and explicit cloud consent plus pixel-review confirmation. Show operator-configured cost estimates and enforce requested estimated-spend limits.
• Add configured loopback vision inference, provider SDK installation extras and the current Google Gen AI SDK.
• Add analysis provenance, lexical reference-report comparison, preliminary FHIR R4/DICOM Basic Text SR exports and saved-report evaluation with baseline regression checks.
• Extend provider-install/browser CI, Docker dependency updates and persistent storage configuration; update CodeQL pins to 4.38.0 and document operation, retention and research limitations.
• Preserve the exact vendored htmx bytes in pre-commit formatting so its Subresource Integrity hash remains valid.

Testing

• Python suite: 338 passed, 1 skipped; coverage 92.10%, above the unchanged 85% minimum.
• New pipeline, privacy, provider, evaluation, export and web integration tests.
• Real Chromium user journeys on desktop and mobile using synthetic DICOM: upload, study selection, local job, viewer, downloads, API-key access, refresh/reconnect and deletion. Deterministic browser mocks cover additional cancellation, network-failure and finding-review states; backend integration tests separately cover actual cancellation and review persistence.
• Ruff lint and formatting, Mypy, Bandit, JavaScript syntax and YAML parsing.
• All pre-commit hooks pass.
• Isolated cloud SDK installation/import smoke test, without external inference calls.

The optional real-Torch batch-invariance test was skipped locally because Torch is absent. Docker image builds, hosted CI, live cloud/local vision-model inference and clinical evaluation were not run locally.

Additional Notes

• This remains a single-process, single-user research workbench, not a validated medical device. An API key does not provide tenant isolation or professional identity verification.
• Browser local mode uses statistics without model downloads and generates no diagnostic findings. Vision inference requires separately configured models/providers.
• Cloud analysis now requires explicit pixel-review confirmation (--pixels-reviewed in the CLI) in addition to transmission consent. Metadata scrubbing/OCR do not guarantee anonymous pixels or free text.
• Cost estimates are operator-supplied and are not hard provider billing caps. Report comparison is lexical; exports remain preliminary/unverified and need receiving-system validation.
• Original uploads, normalized images, reports and review history persist in MEDCHECK_STATE_DIR; users must explicitly delete retained data. Completed results survive restart; interrupted jobs must be rerun.

See docs/workbench.md, docs/evaluation.md and CHANGELOG.md for usage, configuration and limitations.

Summary by CodeRabbit

• New Features

  • Added a complete browser workbench for DICOM/ZIP upload, study selection, queued analysis, progress tracking, cancellation, resumption, and deletion.
  • Added slice viewing, quality warnings, finding review and audit history, image references, and downloadable JSON, HTML, PDF, FHIR, and DICOM-SR reports.
  • Added de-identification, OCR/manual pixel redaction, reference-report comparison, and local loopback vision-server support.
  • Added DICOMDIR and single-file imports, study selection, evaluation tooling, and configurable storage and processing limits.

• Bug Fixes

  • Improved DICOM ordering, decoding, validation, report safety, provider detection, and persistence across restarts.

• Documentation

  • Updated setup, workbench, model, evaluation, intended-use, and limitations guidance.

[coderabbitai]

Navigate logical layers of code changes, visualize relationships, and explore their blast radius.

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: 7149bbe5-23a7-4ce5-ad15-7591480a75b6

✨ Finishing Touches

📝 Generate docstrings

• Commit to this branch
• Create a new PR

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[codecov]

Codecov Report

❌ Patch coverage is 93.84089% with 72 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/medcheck/web/jobs.py	92.71%	8 Missing and 14 partials ⚠️
src/medcheck/providers/local.py	77.55%	5 Missing and 6 partials ⚠️
src/medcheck/pipeline/preprocess.py	87.50%	5 Missing and 3 partials ⚠️
src/medcheck/web/api.py	94.73%	7 Missing and 1 partial ⚠️
src/medcheck/pipeline/ml_analysis.py	75.86%	4 Missing and 3 partials ⚠️
src/medcheck/pipeline/vision_analysis.py	81.81%	3 Missing and 3 partials ⚠️
src/medcheck/main.py	83.33%	4 Missing and 1 partial ⚠️
src/medcheck/llm/local.py	93.87%	1 Missing and 2 partials ⚠️
src/medcheck/pipeline/privacy.py	99.38%	0 Missing and 1 partial ⚠️
src/medcheck/web/app.py	92.30%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!