gliostitch

Reproducible cohort selection over four public glioma MRI datasets — stitched into one manifest of facts, with a study definition layered on top.

Assembling a multi-cohort GBM study means reconciling four different directory layouts, four clinical CSV conventions, three ID formats, two segmentation label schemes, and an unknown number of patients who appear in more than one dataset. Most of that work gets done once, in a notebook, and is difficult to reproduce six months later.

This project does it as a pipeline. It emits one manifest of facts about every imaging session, and a separate layer that turns those facts into a cohort according to a named, versioned study definition. The split matters: the manifest never changes when the research question does.

1661 imaging sessions  →  study definition  →  502 sessions (377 training / 125 external)
                                                with a reason attached to all 1159 exclusions

Datasets: BraTS-2020 · UCSF-PDGM · UPENN-GBM · RHUH-GBM (not redistributed — you supply the data and configure its location)

Install

pip install gliostitch              # selection + path resolution
pip install "gliostitch[pipeline]"  # + build the manifest yourself
pip install "gliostitch[load]"      # + read NIfTI volumes
pip install "gliostitch[monai]"     # + framework backends: monai / torchio / torch

The core is pandas, numpy and pydantic only. If you just want to query a manifest someone else built, you never install the pipeline.

Quickstart

Build the manifest

Point config/pipeline.yaml at your dataset roots, then:

gliostitch verify-layout   # check the directories before doing any work
gliostitch build           # → output/master_manifest.csv

Select a cohort

from gbm_os import Cohort
from gbm_os.studies import GBM_OS_STUDY

cohort = Cohort.from_manifest(
    "output/master_manifest.csv",
    data_roots={"brats2020": "/data/BraTS-2020", ...},
)

view = GBM_OS_STUDY.apply(cohort)
print(len(view))                      # 502

for spec in view:                     # SampleSpec: resolved paths + clinical metadata
    print(spec.paths["t1ce"], spec.os_days, spec.os_class)

Or express criteria directly — all optional, all composable:

view = cohort.select(
    datasets=["ucsf_pdgm", "rhuh_gbm"],
    baseline_only=True,
    require_complete=True,
    filters={"eor": "GTR", "who_grade": [4, None]},
    where=lambda r: r["age"] is not None and r["age"] >= 18,
    resolve_duplicates="drop",
)

Load it

ds = view.to_monai()          # or .to_torchio() / .to_torch()
folds = view.split(k=5, seed=42)      # stratified, patient-grouped, deterministic
train = folds.fold(0, "train")

Why it is built this way

The manifest records facts, not decisions. One row per imaging session, 31 columns, covering identity, paths, modality presence, and normalised clinical values. It does not store os_class, is_baseline or partition — those are modelling choices, derived at load time. A fact discarded at ingest is unrecoverable; a derivation recomputed on load costs milliseconds.

A study is a separate, versioned object. Eligibility criteria, survival thresholds and the held-out cohort live in gbm_os.studies, declaratively — no lambdas, so a definition can be printed, diffed and pasted into a methods section.

$ gliostitch studies gbm-os
gbm-os v2
Overall-survival classification over baseline preoperative GBM MRI, held out on UPENN-GBM.

Eligibility:
  - baseline session only (session_index == 0)
  - all four structural modalities present
  - eor == 'GTR'
  - who_grade in [4, None]
  - has_os == True

Changing the study cannot change the manifest — building under a different one produces a byte-identical CSV. That property is enforced by a test.

Every exclusion is accounted for. Selection records what each criterion removed, so a cohort table is derivable rather than asserted:

>>> print(view.provenance())
input                      1661
baseline_only              1661 →   1521  (−140)
require_complete           1521 →   1126  (−395)
filter:eor                 1126 →    526  (−600)
filter:who_grade            526 →    509  (−17)
filter:has_os               509 →    503  (−6)
resolve_duplicates          503 →    502  (−1)
selected                    502

>>> view.exclusions()      # every dropped session, tagged with its reason

len(view) + len(view.exclusions()) always equals the input.

Duplicates are annotated, never deleted. Patients appearing in more than one dataset are flagged with a group id and a confidence level. Which copy to keep is a selection policy, not a manifest edit.

Documentation

Development

source activate.sh
pytest                   # 246 tests
pytest tests/manifest    # pipeline only — runs against a synthetic dataset tree,
                         # no real data needed, ~1 s

The pipeline suite builds a miniature four-dataset tree on the fly and runs the real adapters against it, so join contracts are testable without access to the MRI data. Tests that need a built manifest skip themselves with a reason.

Status

Beta. The cohort figures above are current as of the latest entry in CHANGELOG.md, which records data-affecting changes separately from code changes — a fix that moves a cohort count is breaking for anyone who published against the old one.

Citation

If you use this in published work, please cite it. GitHub's Cite this repository button reads CITATION.cff and will produce BibTeX or APA for you.

Lim, Wei Xin (Davian). gliostitch: reproducible cohort selection over public glioma MRI datasets, version 0.1.0, 2026. https://github.com/realdavian/gliostitch

ORCID: 0009-0005-2683-1528

Please also cite the source datasets — BraTS-2020, UCSF-PDGM, UPENN-GBM and RHUH-GBM each have their own citation requirements, and this project redistributes none of them.

Licence

GNU General Public License v3.0 or later.

You may use, modify and redistribute this software, including commercially. If you distribute a modified version, or software that incorporates this one, that work must also be released under the GPL. Running it to produce research results places no obligation on your results — only on distributed software.