CheXpert-Plus¶

Modality: CXR | Format: DICOM | Dim: 2D | Labels: 14 pathologies + reports

Overview¶

CheXpert-Plus is an extension of the original CheXpert dataset that pairs de-identified radiology reports (and patient demographics) with 223,228 chest X-rays across 187,711 studies from 64,725 patients. The images are pixel-identical to CheXpert v1.0 — 186 studies from the original release are absent because their reports could not be recovered — and the release additionally ships DICOM and PNG image formats.

Download¶

Available from the Stanford AIMI Shared Datasets page. Requires registration.

The metadata CSV's path_to_dcm column stores paths like train/patient42142/study5/view1_frontal.dcm, so base_image_dir must be the directory that has train/ and valid/ as immediate subdirectories — i.e. DICOM/Uncompressed/ for the official DICOM release.

Expected layout¶

chexpertplus/
  df_chexpert_plus_240401.csv
  findings_fixed.json
  impression_fixed.json
  report_fixed.json                 # de-identified reports (optional)
  DICOM/
    Uncompressed/                   # ← base_image_dir points here
      train/
        patient00001/study1/view1_frontal.dcm
        ...
      valid/
        ...
    Compressed/
      ...
  PNG/
    png_chexpert_plus_chunk_*.zip   # PNG ships as zipped chunks

Label columns¶

Same 14 labels as CheXpert: atelectasis, cardiomegaly, consolidation, edema, enlarged_cardiomediastinum, fracture, lung_lesion, lung_opacity, no_finding, pleural_effusion, pleural_other, pneumonia, pneumothorax, support_devices

Constructor arguments¶

Argument	Type	Required	Default	Description
`base_image_dir`	`str`	Yes	`None`	Direct parent of `train/` and `valid/` (the `path_to_dcm` column starts with `train/...` or `valid/...`). For the canonical DICOM release: `/data/chexpertplus/DICOM/Uncompressed/`
`csv_path`	`str`	No	auto	`df_chexpert_plus_240401.csv`; auto-discovered
`label_json_path`	`str`	No	auto	`report_fixed.json`; auto-discovered near `base_image_dir`
`drop_uncertain`	`bool`	No	`True`	Drop rows with uncertain labels

Shared arguments (inherited from `BaseRadiologicalDataset`)¶

Argument	Type	Required	Default	Description
`output_cls`	`bool`	No	`False`	Include `"cls"` in data dict
`output_mask`	`bool`	No	`False`	Not supported; ignored
`output_report`	`bool`	No	`False`	Include `"report"` (text) in data dict
`output_bbox`	`bool`	No	`False`	Not supported; ignored
`transform`	MONAI transform	No	`None`	MONAI Compose transform; `None` uses the default pipeline
`cache_dir`	`str`	No	`"./cache"`	MONAI 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 saved harmonized CSV

Dataset constructor¶

import torch
from radharmony.dataset import CheXpertPlusDataset

ds = CheXpertPlusDataset(
    base_image_dir="/data/chexpertplus/DICOM/Uncompressed/",   # contains train/ and valid/
    csv_path="/data/chexpertplus/df_chexpert_plus_240401.csv",
    label_json_path="/data/chexpertplus/report_fixed.json",
    output_cls=True,
    output_report=True,
    dtype=torch.float32,
)
train_ds, val_ds = ds.get_datasets(n_splits=5)

Harmonizer¶

from radharmony.harmonizer import CheXpertPlusHarmonizer

h = CheXpertPlusHarmonizer(
    csv_path="/data/chexpertplus/df_chexpert_plus_240401.csv",
    base_image_dir="/data/chexpertplus/DICOM/Uncompressed/",
    label_json_path="/data/chexpertplus/report_fixed.json",
)
df = h.harmonize()
print(df.columns.tolist())
df.to_csv("chexpert_plus_harmonized.csv", index=False)

Load from saved harmonized CSV¶

import pandas as pd
from radharmony.dataset import CheXpertPlusDataset

ds = CheXpertPlusDataset(
    base_image_dir="/data/chexpertplus/DICOM/Uncompressed/",
    harmonized_df=pd.read_csv("chexpert_plus_harmonized.csv"),
    output_cls=True,
    output_report=True,
)

Harmonizer notes¶

Reads the main CSV and optionally report_fixed.json for de-identified report text
Uncertain labels are handled the same way as CheXpert (drop_uncertain)
Bounding boxes and masks (when present) are normalised to fractional coordinates

Outputs¶

Flag	Key	Shape	Notes
`output_cls=True`	`"cls"`	`(14,)`	Binary labels
`output_report=True`	`"report"`	`str`	De-identified report text

Example paths¶

Role	Path
`base_image_dir`	`/path/to/CheXpert_Plus/chexpertplus/DICOM/Uncompressed`
`csv_path` (`df_chexpert_plus_240401.csv`)	`/path/to/CheXpert_Plus/chexpertplus/df_chexpert_plus_240401.csv`
`label_json_path` (`report_fixed.json`)	`/path/to/CheXpert_Plus/chexpertplus/report_fixed.json`

impression_fixed.json and findings_fixed.json also exist alongside report_fixed.json. CSVs and JSONs live 2 levels above base_image_dir and are auto-discoverable via infer_path.