Spaces:
Running
Running
Download pkg/src/strokecopilot/data/isles24.py from parsa-hke/stroke-copilot: direct link, hf CLI and curl.
- Browser
- Download file 4.15 kB
-
https://huggingface.co/spaces/parsa-hke/stroke-copilot/resolve/main/pkg/src/strokecopilot/data/isles24.py
- Command line
-
hf download hf://spaces/parsa-hke/stroke-copilot/pkg/src/strokecopilot/data/isles24.py
-
curl -L -o isles24.py https://huggingface.co/spaces/parsa-hke/stroke-copilot/resolve/main/pkg/src/strokecopilot/data/isles24.py
4.15 kB
| """ISLES'24 slim-format reader: list_cases(), load_case(), fetch_slim() (plan §5.4). | |
| Held-out cases are refused before Phase 4, and list_cases() leaves them out. Arrays are RAS, on one | |
| isotropic grid per manifest.json, except ncct_native. Tmax voxels at -30 have no perfusion data, and | |
| map background comes from the manifest. Reference volumes are in `native_volumes`, measured before | |
| resampling, because resampling moves Tmax > 6 s by up to about 10 mL (docs/DATA.md). | |
| """ | |
| import json | |
| import os | |
| import subprocess | |
| from dataclasses import dataclass, field | |
| from functools import lru_cache | |
| import nibabel as nib | |
| import numpy as np | |
| import pandas as pd | |
| from strokecopilot.config import LOCAL_ROOT | |
| from strokecopilot.data import clinical_columns, splits | |
| from strokecopilot.data.archive_listing import copy_with_progress | |
| SLIM_NAME = "isles24_slim_v1" | |
| DEFAULT_ROOT = LOCAL_ROOT / SLIM_NAME | |
| FILES = { | |
| "ncct": "ncct.nii.gz", | |
| "cbf": "cbf.nii.gz", | |
| "cbv": "cbv.nii.gz", | |
| "mtt": "mtt.nii.gz", | |
| "tmax": "tmax.nii.gz", | |
| "occlusion": "cta_occlusion.nii.gz", | |
| "cow": "cow.nii.gz", | |
| "infarct": "infarct_followup.nii.gz", | |
| } | |
| MASKS = ("occlusion", "infarct") | |
| class Case: | |
| id: str | |
| arrays: dict | |
| spacing_mm: float | |
| affine: np.ndarray | |
| clinical: dict | |
| native_volumes: dict = field(default_factory=dict) | |
| def note_fields(self): | |
| """Clinical values allowed in a triage note (plan §8.3). `clinical` also holds forbidden ones.""" | |
| return {column: value for column, value in self.clinical.items() if column != "case" and clinical_columns.tag_of(column) == "allowed"} | |
| def fetch_slim(tar_path, parent=LOCAL_ROOT): | |
| """Copy the tarball to local disk and unpack it (DEC-3). Does nothing if already there.""" | |
| root = os.path.join(parent, SLIM_NAME) | |
| if os.path.exists(os.path.join(root, "manifest.json")): | |
| return root | |
| os.makedirs(parent, exist_ok=True) | |
| local = os.path.join(parent, f"{SLIM_NAME}.tar") | |
| copy_with_progress(tar_path, local) | |
| subprocess.run(["tar", "-xf", local, "-C", parent], check=True) | |
| os.remove(local) | |
| return root | |
| def load_manifest(root=DEFAULT_ROOT): | |
| with open(os.path.join(root, "manifest.json")) as f: | |
| return json.load(f) | |
| def list_cases(root=DEFAULT_ROOT, folds=None): | |
| """Cases in the tarball, without held-out cases unless ALLOW_HELDOUT=1.""" | |
| cases = load_manifest(root)["cases"] | |
| if splits.heldout_allowed(): | |
| return cases | |
| heldout = set((folds or splits.load_folds())["heldout"]) | |
| return [case for case in cases if case not in heldout] | |
| def _tables(root): | |
| clinical = pd.read_csv(os.path.join(root, "clinical.csv"), dtype=object).set_index("case") | |
| volumes = pd.read_csv(os.path.join(root, "native_volumes.csv")).set_index("case") | |
| return clinical, volumes | |
| def _clean(row): | |
| return {column: (None if pd.isna(value) else value) for column, value in row.items()} | |
| def load_case(case, root=DEFAULT_ROOT, names=tuple(FILES), folds=None): | |
| """Arrays, clinical row, spacing and native volumes for one case. Raises on a held-out case.""" | |
| splits.check_ids([case], folds) | |
| root = str(root) | |
| if case not in load_manifest(root)["cases"]: | |
| raise KeyError(f"Case {case} is not in the tarball.") | |
| arrays, affine = {}, None | |
| for name in names: | |
| img = nib.load(os.path.join(root, case, FILES[name])) | |
| array = np.asanyarray(img.dataobj) | |
| arrays[name] = array > 0 if name in MASKS else array | |
| affine = img.affine | |
| clinical, volumes = _tables(root) | |
| return Case( | |
| id=case, | |
| arrays=arrays, | |
| spacing_mm=load_manifest(root)["spacing_mm"], | |
| affine=affine, | |
| clinical={"case": case, **_clean(clinical.loc[case])}, | |
| native_volumes={k: float(v) for k, v in volumes.loc[case].items()}, | |
| ) | |
| def load_native_ncct(case, root=DEFAULT_ROOT, folds=None): | |
| """The unchanged source NCCT, as a nibabel image, for Vision B (plan §8.2). Not RAS or resampled.""" | |
| splits.check_ids([case], folds) | |
| return nib.load(os.path.join(str(root), case, "ncct_native.nii.gz")) | |