Liver Working Tree

The drive layout behind the liver tumour-origin project: a git repository buried one level down where all the real work lives, twenty cloned model environments, and a documentation site that deploys from a third repository.

Purpose

This page is about where things are, not what they conclude — the science is on Liver Tru-Cut — Tumour Origin AI. It exists because the layout is genuinely confusing and cost time to work out: K:\liver itself is not a git repository, but K:\liver\dataPreperation is, and that subfolder is where essentially everything lives. Anyone told “the liver project is on K:” and looking at the top level sees virtualenvs and data folders and concludes there is no code.

Data used

extractedData/ holds the 509 pseudonymised cases of the main cohort as extracted SVS files; rawData/ holds the corresponding Sectra ZIP exports. The raw count appears larger than it is because a Mac-to-Windows copy left macOS ._* resource-fork sidecars alongside the real archives — worth knowing before anyone counts files to estimate the cohort.

Two loose files at the top level are the clinical ground truth: liver_tru_cut_tumor.fmp12 and liver_tru_cut_tumor.xlsx. This is a different database from the one on Liver Tru-Cut Cohort, which belongs to the small four-case morphometry set. Two liver cohorts, two FileMaker databases, similar names — the single most likely thing to get wrong here.

Methods

What each top-level directory is, read rather than inferred:

Directory	What it is
`dataPreperation/`	The project. Its own git repo: a Quarto book (`LIVER_GUIDE.qmd` at 348 KB, `pipeline.qmd` as a runbook), ~90 Python scripts, Groovy QuPath scripts, and all `outputs/` — manifests, features, results, EDA reports
`hpl/`	Histomorphological Phenotype Learning, adapted from Quiros et al. — the model-comparison engine that produced the cross-scanner leaderboard
`repos/`	Cloned upstream model libraries used as import-time dependencies: CONCH, MADELEINE, MUSK, prov-gigapath
`livertumorweb/`	Not code — the git-tracked deploy target (`sbalci/livertumorweb`) for the rendered Quarto site, mirrored from `dataPreperation` by a post-render hook
`extractedData/`, `rawData/`, `dataPreperation/`	The cohort, above

The seven virtualenvs turned out to mean exactly what they looked like. venv_conch, venv_gigapath, venv_grandqc, venv_medgemma, venv_musk and venv_chief each back a specific inference script, and venv_livertumor is the main pipeline environment rather than a model. An earlier version of this page could only say that environments existed and marked their use [unverified]; that is now resolved — every one except CHIEF has a corresponding results file. See Liver Tru-Cut — Tumour Origin AI for what they produced.

livertumorweb is the third instance of a pattern now worth naming. A project writes in one repository, renders, and mirrors the output into a second repository that exists only to publish — exactly what PinkKidneyWeb does for PembeBobrek. Both mirrors have an empty or near-empty README by design. It works, and it reliably confuses anyone who finds the published page first and goes looking for its source in the wrong place.

Current state / open questions

Active. This is where the work described on Liver Tru-Cut — Tumour Origin AI is actually done.

The top-level directory is not under version control, and it holds the cohort, the clinical databases and the model environments. dataPreperation/ being a repository protects the code and the outputs; nothing protects the rest. The clinical FileMaker database is the one irreplaceable artefact — slides can be re-exported, a hand-curated label table cannot.
repos/ duplicates model code that is also on the Foundation Model Cache drive, and neither location records a pinned revision. Two copies of an upstream library on two drives, both at unknown commits, is the reproducibility gap that page already flags, now with a second instance.
pip-builds/ is leftover pip and wandb temporary output, not project material — safe to ignore, and worth deleting if the drive ever needs space.

Related: Liver Tru-Cut — Tumour Origin AI — the science this directory produces; read that first. Related: Liver Tru-Cut Cohort — the other liver folder and the other FileMaker database. Related: Foundation Model Cache — the model weights these environments consume, plus the duplicated clones in repos/. Related: PinkKidneyWeb — the same render-and-mirror publishing pattern, in the kidney project. Related: hepatocyte-app — the sibling pipeline working on the smaller liver set.

Derived from: directory reading 2026-07-29 — top-level listing of K:/liver, dataPreperation/ structure and its CLAUDE.md/README.md/ABOUT.md, hpl/README.md, livertumorweb/ git remote and README, repos/ contents, and confirmation that no .git exists at K:/liver itself while one exists at K:/liver/dataPreperation.