MekaNet

A manuscript-aligned public release for megakaryocyte detection and myeloproliferative neoplasm classification

Important

This repository is intentionally trimmed for manuscript review and reproducibility. It distinguishes between: what the paper describes, what is publicly shipped here, and what remains private.

At A Glance

Item	Public status	Source
Manuscript	Not public	Removed from the public release
Detection weight	Public	HF model repo
Public tabular feature table	Not public	Removed from the public release
Patient-level master cohort	Not public	Not included in this release
Private image manifests and full clinical tables	Not public	Not included in this release

Quick Links

• Citation metadata: CITATION.cff
• Alignment note: docs/manuscript_alignment.md
• Weight policy: weights/README.md
• Data policy: data/README.md

Manuscript Alignment

• The private manuscript uses Department of Laboratory Medicine, not Department of Pathology.
• Public-facing docs in this repository were rewritten to match manuscript wording and remove stale release-summary material.
• The repository now prioritizes auditability over promotional presentation.

Public Data Accounting

The current repository supports the following components with local evidence:

Scope	Public evidence in repo	Notes
Detection training cohort	Manuscript describes a partially labeled B-hospital MPN image subset	Image-level manifest is not public
Detection internal fully labeled evaluation	Manuscript describes fully labeled images used for internal testing	Image files are not public
External validation	Manuscript describes S-hospital test sets used for generalizability	Detailed in the manuscript evaluation

The following items are not present in this public release to maintain anonymity and compliance:

• The complete patient-level raw dataset
• The full patient/control manifest used to reconstruct the exact classification cohort from raw IDs
• The control cases cited in the classification section of the manuscript
• Detailed ID mappings for hospital-specific subsets
• Historically uploaded public tabular feature tables
• Production model weights beyond the released detection checkpoint

For a detailed reconciliation note, see docs/manuscript_alignment.md.

Repository Layout

.
├── CITATION.cff
├── README.md
├── data/
│   └── README.md
├── experiments/
│   ├── classification/
│   └── detection/
├── mekanet/
├── weights/
└── run_paper_reproduction.py

Module Roles

• mekanet/: reusable Python package components
• experiments/detection/: manuscript-facing detection pipeline
• experiments/classification/: manuscript-facing classification analysis
• data/: documentation about non-public data scope
• weights/: weight handling policy and verification helpers

Canonical Entry Points

For this slimmed public release, the canonical reproducibility entry points are:

python run_paper_reproduction.py --quick --dry-run
python experiments/classification/run_all_experiments.py

Legacy deployment wrappers, benchmark helpers, auto-generated reports, unused config variants, and unreferenced figures were removed to keep the repository reviewer-friendly.

Current Execution Status

The codebase has been bootstrapped so that package imports and dry-run entrypoints work in a clean local environment after installing dependencies.

pip install -r requirements.txt
python run_paper_reproduction.py --quick --dry-run
python experiments/classification/run_all_experiments.py

Current limitations

• Detection execution still requires the released weight artifact.
• Classification execution still requires a user-supplied data/demo_data/classification.csv.
• No public tabular feature CSV is shipped in this repository.

Public Assets

Detection Weight

• repo: LeBrony/mekanet-release-weights
• direct file: epoch60.pt

The local downloader uses the Hugging Face source first and verifies the SHA-256 recorded in weights/manifest.json.

Reviewer Notes

Note

Earlier drafts of this repository contained release-summary, quickstart, benchmark, and deployment files with stale cohort counts or non-canonical execution paths. Those were removed to reduce ambiguity during review.

The remaining documentation favors manuscript-aligned descriptions over broad showcase material.

The manuscript source file itself is private and is not distributed in this repository.

Clinical Affiliation

• Sang Mee Hwang: Department of Laboratory Medicine, Seoul National University Bundang Hospital
• Young-eun Lee: Department of Laboratory Medicine, Seoul National University College of Medicine

License

License scope is intentionally split:

• code in this GitHub repository: MIT
• public weight artifacts: separate review/research-only terms

See:

• LICENSE
• weights/WEIGHTS_LICENSE.md

Patient-level source data is not included in this public release.