Skip to content

ReXGradient-160K

Modality: CXR | Format: PNG | Dim: 2D | Labels: none | Reports: English (4-section)

Overview

ReXGradient-160K is the largest publicly available multi-site chest X-ray dataset: 273,004 PNG images from 160,000 studies / 109,487 patients across 3 U.S. health systems (79 medical sites), distributed via HuggingFace by Rajpurkar Lab. Images were downsampled to 25 % of their original dimensions with cubic interpolation + anti-aliasing under the data-use agreement; the original-resolution images are commercially available through Gradient Health.

The dataset ships no structured pathology labels — only free-text reports — and pre-segments each report into four sections: Indication, Comparison, Findings, Impression. The harmonizer concatenates them with section headers into a single report string per image (one report shared across all images of the same study).

RadHarmony provides three split-specific classes: - ReXGradientTrainDataset — 238,968 images from 140,000 studies / 95,716 patients - ReXGradientValidDataset — 17,007 images from 10,000 studies / 6,964 patients - ReXGradientTestDataset — 17,029 images from 10,000 studies / 6,807 patients (public test)

An additional private test set of 10,000 studies is reserved for the ReXrank benchmark and is not distributed here.

Download

Primary source: HuggingFace dataset rajpurkarlab/ReXGradient-160K. Access is gated under the ReXGradient-160K Non-Commercial Data Access and Use Agreement (Harvard Medical School / Gradient Health) — non-commercial research only, no clinical use.

After accepting the agreement on HuggingFace:

hf download rajpurkarlab/ReXGradient-160K \
    --repo-type dataset \
    --local-dir /data/ReXGradient-160K/download

The PNGs ship as 10 concatenated zstd-compressed tar parts (deid_png.part00 … deid_png.part09, ~15 GB each). Concatenate + extract:

cd /data/ReXGradient-160K
cat download/deid_png.part* | tar --zstd -xvf - -C .
# → produces ./deid_png/<PatientID>/<AccessionNumber>/studies/<StudyUID>/...

After extraction the part files can be removed (tar returns 0 only on a clean extract). Total extracted size is ~145 GB.

Expected layout

ReXGradient-160K/
  deid_png/                                       ← base_image_dir
    <PatientID>/                                  e.g. GRDNLZHK1CJMB9DS/
      <AccessionNumber>/                          e.g. GRDNLD4ATLU63FN8/
        studies/
          <StudyInstanceUid>/
            series/
              <SeriesInstanceUid>/
                instances/
                  <SopInstanceUid>.png
  download/
    metadata/
      train_metadata_view_position.json           ← csv_path (per split)
      valid_metadata_view_position.json
      test_metadata_view_position.json
      train_metadata.csv                          (per-study, not used by harmonizer)
      valid_metadata.csv
      test_metadata.csv
      train_metadata.json
      valid_metadata.json
      test_metadata.json
      interstitial_pattern_bbox.json              (~398 boxes; not currently wired)
    README.md
    LICENSE

base_image_dir is the deepest stable directory — deid_png/ — whose immediate children are <PatientID>/ directories. The harmonizer's image_path is the rest of the nested path under base_image_dir.

Source of truth: JSON, not CSV

Two parallel metadata forms ship per split:

File Granularity Image paths?
<split>_metadata.csv 1 row per study no
<split>_metadata_view_position.json 1 entry per study with an ImagePath list yes

The CSV alone cannot produce per-image rows — every study has one or more images and the CSV is study-keyed. The harmonizer therefore reads the JSON. Passing a .csv to csv_path raises ValueError with a message pointing at the right JSON filename.

Label columns

None. ReXGradient ships no CheXpert-style structured labels. LABEL_COLS is empty and output_cls=True is not supported (the base class will emit a UserWarning). Deriving labels from the reports (e.g. via CheXbert) is left to the user.

Constructor arguments

Argument Type Required Default Description
base_image_dir str Yes None Root of the extracted PNG tree, e.g. /data/ReXGradient-160K/deid_png/. Immediate children are <PatientID>/ directories.
csv_path str No auto Path to <split>_metadata_view_position.json. Must be the JSON, not the CSV. Auto-discovered near sibling/ancestor metadata/ directories.

Shared arguments (inherited from BaseRadiologicalDataset)

Argument Type Required Default Description
output_cls bool No False Not supported — dataset has no structured labels
output_mask bool No False Not supported
output_report bool No False Include "report" (4-section concatenated text) in the data dict
output_bbox bool No False Not supported by this dataset class (a separate interstitial-pattern bbox JSON exists for ~398 examples but is out of scope here)
transform MONAI transform No None MONAI Compose transform; None uses the default pipeline
cache_dir str No "./cache" MONAI PersistentDataset cache directory
dtype torch.dtype No torch.bfloat16 Output tensor dtype
harmonized_df pd.DataFrame No None Pre-built harmonized DataFrame
harmonizer harmonizer No None Pre-instantiated harmonizer
harmonizer_path str No None Path to a saved harmonizer pickle

Dataset constructor: train

import torch
from radharmony.dataset import ReXGradientTrainDataset

ds = ReXGradientTrainDataset(
    base_image_dir="/data/ReXGradient-160K/deid_png/",
    output_report=True,
    dtype=torch.float32,
)
train_ds, val_ds = ds.get_datasets(n_splits=10)
sample = train_ds[0]
# sample["img"]    → torch.Tensor, shape (1, 224, 224), float in [-1, 1]
# sample["report"] → "INDICATION:\n...\n\nCOMPARISON:\n...\n\nFINDINGS:\n...\n\nIMPRESSION:\n..."

Dataset constructor: validation

from radharmony.dataset import ReXGradientValidDataset

ds = ReXGradientValidDataset(
    base_image_dir="/data/ReXGradient-160K/deid_png/",
    output_report=True,
)
full_ds = ds.get_datasets()   # 17,007 samples

Dataset constructor: public test

from radharmony.dataset import ReXGradientTestDataset

ds = ReXGradientTestDataset(
    base_image_dir="/data/ReXGradient-160K/deid_png/",
    output_report=True,
)
test_ds = ds.get_datasets()   # 17,029 samples

Harmonizer

from radharmony.harmonizer import ReXGradientValidHarmonizer

h = ReXGradientValidHarmonizer(
    csv_path="/data/ReXGradient-160K/download/metadata/valid_metadata_view_position.json",
    base_image_dir="/data/ReXGradient-160K/deid_png/",
)
df = h.harmonize()
print(df.columns.tolist())
# ['patient_id', 'study_id', 'image_path', 'view_position', 'report',
#  'accession_number', 'patient_sex', 'patient_age', 'study_date']
df.to_csv("rexgradient_valid_harmonized.csv", index=False)

Load from saved harmonized CSV

import pandas as pd
from radharmony.dataset import ReXGradientValidDataset

ds = ReXGradientValidDataset(
    base_image_dir="/data/ReXGradient-160K/deid_png/",
    harmonized_df=pd.read_csv("rexgradient_valid_harmonized.csv"),
    output_report=True,
)

Harmonizer notes

  • JSON source, per-image rows: each JSON entry is one study with parallel lists (ImagePath, ImageViewPosition, ImageShape, ImageModality, ImageBodyPart). The harmonizer explodes by ImagePath so that each row in the harmonized DataFrame corresponds to one image. The 4 report sections (Indication / Comparison / Findings / Impression) are duplicated across all images of the same study.
  • Path stripping: JSON ImagePath entries are written relative to metadata/ as "../deid_png/<...>"; the harmonizer strips the leading "../deid_png/" so the result is relative to base_image_dir.
  • csv_path argument: kept for interface consistency with the rest of RadHarmony, but it must point at the per-image JSON. A .csv path is rejected with an explicit error.
  • View position: read inline from ImageViewPosition (one entry per image). Values include PA, AP, LATERAL, LL, POSTERO_ANTERIOR, LAO, RAO, UNKNOWN, N/A, and ~2 % None. No DICOM header reads required.
  • Reports: combined with section headers via _format_report(indication, comparison, findings, impression). Empty / "None" / NaN sections are skipped; rows with all four sections empty get a None report.
  • Extra metadata columns: accession_number, patient_sex, patient_age (raw string, e.g. "007Y" / "002D"), study_date (YYYYMMDD int) are exposed via EXTRA_OUTPUT_COLS for downstream filtering.
  • Bboxes: metadata/interstitial_pattern_bbox.json contains ~398 Label-Studio-format boxes (x / y / width / height in original pixel coords, label = subtype of interstitial pattern). This v1 integration does not wire them in. To add them later, normalize coordinates by the per-image ImageShape and join on the synthetic <study_id>_<sop_uid>.png key the bbox file uses.

Outputs

Flag Key Shape Notes
output_report=True "report" str 4-section concatenated text (INDICATION / COMPARISON / FINDINGS / IMPRESSION). None when all 4 source sections are empty.
output_cls=True — — Not supported; no structured labels
output_mask=True — — Not supported
output_bbox=True — — Not supported (see Harmonizer notes for the interstitial-pattern bbox file)