longlephamtien/OSIRIS_Data

Lumbar spine MRI -- ontology-mapped, per-patient restructured

Raw MRI files from each source dataset, reorganized into one flat folder per patient with a single naming pattern shared across every source, plus the ontology-mapped labels and an audit trail back to the original raw files. See metadata/ontology.yaml for the condition/severity knowledge base every condition_id/scale_id below resolves against.

Repo layout

README.md
dataset_info.json          machine-readable counterpart to this file
extract.py                 cross-platform (Mac/Windows/Linux) decompressor, stdlib only
rsna.tar.gz        compressed -- see 'Extracting the data' below
aisslab.tar.gz     compressed -- see 'Extracting the data' below
spider.tar.gz      compressed -- see 'Extracting the data' below
mendeley.tar.gz    compressed -- see 'Extracting the data' below
metadata/
  ontology.yaml
  unify_outputs/rsna_unified.jsonl
  unify_outputs/aisslab_unified.jsonl
  unify_outputs/spider_unified.jsonl
  unify_outputs/mendeley_series_manifest.jsonl
  tokens.jsonl              filtered to included sources
  file_mapping.jsonl        filtered to included sources

Extracting the data

Each source ships as one <source>.tar.gz at repo root instead of thousands of loose files (RSNA alone is 147k+ files) -- download/clone this repo, then from the repo root run:

python extract.py

Standard library only (tarfile) -- no pip install, no system tar/gzip binary required, works unmodified on macOS, Windows, and Linux with whatever Python 3 is already installed. It extracts every *.tar.gz next to it and deletes the archive after a successful extraction (--keep-archives to keep them, --sources rsna,spider to extract only some). Equivalent manual command if you'd rather not run the script: tar xzf <source>.tar.gz.

Directory structure (inside each archive, after extracting)

<source>/<patient_id>/
  <patient_id_tag>__<series_tag>__<role>__<instance>.<ext>   (one file per raw file)
  findings.jsonl                                             (ontology-mapped labels for this patient)

Every dataset produces this exact same shape: one directory level (patient folders), fully flat inside -- no nested series/mask/annotation subfolders. Every identifying field lives in the filename itself, not in folder nesting, so a file stays unambiguous even if copied out of its folder.

Naming pattern

{patient_id_tag}__{series_tag}__{role}__{instance}.{ext}

Component	Meaning	Example
patient_id_tag	{source}-{patient_id} -- embedded even though the folder path already disambiguates source, because two sources are known to have overlapping numeric patient-id ranges for different real people	aisslab-0389
series_tag	{plane_abbrev}-{weighting} (sag/ax/cor/loc/unk); a -2/-3 suffix is added if a patient has 2+ series with the same plane+weighting	sag-T2, ax-T2, sag-T1-2
role	what kind of file this is -- see table below	img, mask
instance	zero-padded 3-digit 1-based slice/instance number, vol for a whole-volume file (.mha), or mid for a single middle slice	007, vol, mid
ext	original file extension, unchanged (no transcoding)	.dcm, .ima, .mha, .xml, .png

Roles present in this export:

role	Meaning
img	primary diagnostic image (DICOM-like instance) or whole volume (.mha)
mask	segmentation mask volume, same geometry as its img sibling
foramina	foraminal-stenosis bounding-box annotation -- .xml (only slices with a box) or .png (every slice)
segxml	vertebra/disc/sacrum polygon segmentation annotation (single middle slice)
segmask	rendered segmentation mask PNG for the segxml polygon above
segmid	plain, unannotated copy of the segmentation middle slice

File field reference

metadata/file_mapping.jsonl (one global file)

One row per physical file in this repo -- the raw<->constructed audit trail.

Field	Meaning
source	dataset name
patient_id	raw, dataset-native patient/study id (not prefixed)
raw_path	path to the original file in the source dataset's own layout
constructed_path	path to this same file's copy in this repo
series_uid	original DICOM SeriesInstanceUID, or null for files with no series identity of their own (e.g. an annotation file)
series_tag	the deduped plane-weighting tag used in this file's constructed name
role	see the roles table above
instance	see the naming-pattern table above

<source>/<patient_id>/findings.jsonl (per patient, ontology-mapped labels)

One JSON object per line, one line per finding -- identical shape to metadata/unify_outputs/<source>_unified.jsonl:

{
  "study_uid": "...", "source": "...",
  "series": [{"series_uid","plane","weighting","n_slices","geometry"}, ...],
  "finding": {
    "condition_id": "...",        // key into ontology.yaml conditions
    "level": "L4_L5",             // vertebral level
    "level_confidence": "anatomical" | "relative_caudal",
    "side": "left" | "right" | "central",
    "severity": {"scale_id": "...", "value": ...},  // scale_id is a key into ontology.yaml scales
    "geometry": {...} | null,     // ANNOTATION geometry (point/box/mask locating the finding) -- not the same as series.geometry (imaging geometry)
    "annotated": true | false      // false = condition in scope but no label for this study -- mask in loss, never a real negative
  }
}

No findings.jsonl for mendeley patients -- it ships with zero labels of any kind (raw PACS export). Its imaging-only inventory is instead in metadata/unify_outputs/mendeley_series_manifest.jsonl: one row per (patient, lumbar study), {study_uid, source, series: [...]} -- same series-ref shape as above, deliberately with no finding key at all.

metadata/tokens.jsonl (filtered to sources in this repo)

One row per finding, rendered into a natural-language [I_p, Y_t] pair for vision-language model training (see tokens/token_utils.py for the templates):

Field	Meaning
study_uid, source	same as findings.jsonl
I_p	context/prompt text: dataset, available imaging series, which condition/level/side is being queried -- never contains the answer
Y_t	the finding/severity text itself -- the training target
meta	condition_id, level, side, scale_id, value, annotated -- the structured fields I_p/Y_t were rendered from

metadata/ontology.yaml

The knowledge base every condition_id/scale_id/level above resolves against: conditions (definition, normal_finding_text, anatomy, external IDs), scales (per-severity-grade clinical descriptions), anatomy, disc_pathology_taxonomy, cooccurrence_and_causal_notes.

dataset_info.json

Machine-readable counterpart to this README -- per-source patient/file/byte counts, citations, and the metadata-file paths above, for a script to read instead of parsing prose.

Why JSONL, not a JSON array

findings.jsonl, metadata/file_mapping.jsonl, metadata/tokens.jsonl and metadata/unify_outputs/*.jsonl are all JSON Lines (one independent JSON object per line) rather than one JSON array, because: (1) streamable -- a training loop or datasets.load_dataset(..., lines=True) reads one record at a time without loading the whole file into memory; (2) appendable -- each row was written incrementally, one per patient/finding, without rewriting the whole file each time; (3) line tools work directly -- wc -l/grep/head all operate correctly with no JSON-aware parser; (4) partial-write safety -- a crash mid-write corrupts at most the last line, not the entire file's syntax the way a truncated JSON array would.

Included sources and citations

• rsna: Richards et al., "The RSNA Lumbar Degenerative Imaging Spine Classification (LumbarDISC) Dataset", Radiology: Artificial Intelligence (2026). https://doi.org/10.1148/ryai.250480
• aisslab: "Medical Spine Sagittal MRI Dataset for Segmentation and Foraminal Stenosis detection", Scientific Data (2026). https://doi.org/10.1038/s41597-026-07138-x
• spider: van der Graaf et al. 2024, "Lumbar spine segmentation in MR images: a dataset and a public benchmark", Scientific Data 11:264. https://doi.org/10.1038/s41597-024-03090-w
• mendeley: Al Kafri, Sudirman et al. 2019, "Boundary Delineation of MRI Images for Lumbar Spinal Stenosis Detection Through Semantic Segmentation Using Deep Neural Networks", IEEE Access 7:43487-43501. https://doi.org/10.1109/ACCESS.2019.2908002

Visibility: private.