MIMIC-CXR-JPG¶

Modality: CXR | Format: JPEG | Dim: 2D | Labels: 14 pathologies

Overview¶

MIMIC-CXR-JPG is the JPEG version of MIMIC-CXR. It contains 227,827 studies (a handful fewer than the DICOM release could not be converted) but stores images as compressed JPEG files, making it faster to load and more suitable for large-scale training. The CheXpert-extracted labels match MIMIC-CXR; the JPG release itself ships no free-text reports.

RadHarmony provides two dataset classes for this dataset: - MIMICCXRJPGDataset — all studies, labelled by the standard CheXpert NLP labeller (mimic-cxr-2.0.0-chexpert.csv). - MIMICCXRJPGTestDataset — strict 687-study labeled test subset, using the independently-relabeled mimic-cxr-2.1.0-test-set-labeled.csv. Useful for fair evaluation since these labels were not derived from the same NLP pipeline used to label the training set.

Download¶

Available at PhysioNet: MIMIC-CXR-JPG. Requires CITI training and signed DUA.

Expected layout¶

mimic-cxr-jpg/2.0.0/
  mimic-cxr-2.0.0-metadata.csv.gz
  mimic-cxr-2.0.0-chexpert.csv             # full-set labels (all studies)
  mimic-cxr-2.1.0-test-set-labeled.csv     # 687-study test set (optional)
  files/
    p10/
      p10000032/
        s50414267/
          02aa804e-bde0afdd-....jpg

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`	`files/` subtree root (e.g. `/data/mimic-cxr-jpg/2.0.0/files/`) — direct parent of the `p10/`, `p11/`, … patient prefix dirs
`csv_path`	`str`	No	auto	`mimic-cxr-2.0.0-metadata.csv.gz`; auto-discovered
`label_csv_path`	`str`	No	auto	`mimic-cxr-2.0.0-chexpert.csv`
`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 — warns "no effect" if `True`
`output_report`	`bool`	No	`False`	Not supported (JPG dataset has no report text) — warns "no effect" if `True`. Use MIMIC-CXR (DICOM) for reports
`output_bbox`	`bool`	No	`False`	Not supported — warns "no effect" if `True`
`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¶

All studies¶

import torch
from radharmony.dataset import MIMICCXRJPGDataset

ds = MIMICCXRJPGDataset(
    base_image_dir="/data/mimic-cxr-jpg/2.0.0/files/",
    output_cls=True,
    dtype=torch.float32,
)
train_ds, val_ds = ds.get_datasets(n_splits=5)

Labeled test subset (687 studies)¶

from radharmony.dataset import MIMICCXRJPGTestDataset

ds_test = MIMICCXRJPGTestDataset(
    base_image_dir="/data/mimic-cxr-jpg/2.0.0/files/",
    output_cls=True,
    dtype=torch.float32,
)
# Auto-discovers mimic-cxr-2.1.0-test-set-labeled.csv near base_image_dir.
# Constructor signature is identical to MIMICCXRJPGDataset; the only
# difference is which label CSV is loaded.

The test variant exposes the same 14 LABEL_COLS as the full dataset. The test CSV's "Airspace Opacity" column is renamed to "Lung Opacity" during preprocessing so the harmonized DataFrame is schema-compatible with the train variant. Joining is on study_id only (the test CSV has no subject_id), so non-test studies are dropped via inner-merge.

Uncertain labels (-1): unlike the train variant, MIMICCXRJPGTestDataset defaults to drop_uncertain=False so every labeled study is retained (test metrics should match the published numbers). -1 values are preserved as-is in the harmonized DataFrame; downstream code can apply U-Zeros (-1 → 0), U-Ones (-1 → 1), or U-Ignore depending on the convention being reproduced. Pass drop_uncertain=True explicitly to drop any study with at least one uncertain label (~226 of 687).

Harmonizer¶

from radharmony.harmonizer import MIMICCXRJPGHarmonizer

h = MIMICCXRJPGHarmonizer(
    csv_path="/data/mimic-cxr-jpg/2.0.0/mimic-cxr-2.0.0-metadata.csv.gz",
    base_image_dir="/data/mimic-cxr-jpg/2.0.0/files/",
    label_csv_path="/data/mimic-cxr-jpg/2.0.0/mimic-cxr-2.0.0-chexpert.csv",
)
df = h.harmonize()
print(df.columns.tolist())
df.to_csv("mimic_cxr_jpg_harmonized.csv", index=False)

Load from saved harmonized CSV¶

import pandas as pd
from radharmony.dataset import MIMICCXRJPGDataset

ds = MIMICCXRJPGDataset(
    base_image_dir="/data/mimic-cxr-jpg/2.0.0/files/",
    harmonized_df=pd.read_csv("mimic_cxr_jpg_harmonized.csv"),
    output_cls=True,
)

Harmonizer notes¶

Primary CSV is the metadata file (mimic-cxr-2.0.0-metadata.csv.gz)
Labels are merged from mimic-cxr-2.0.0-chexpert.csv via study ID
Faster to load than MIMIC-CXR DICOM; preferred for large-scale training

Outputs¶

Flag	Key	Shape	Notes
`output_cls=True`	`"cls"`	`(14,)`	CheXpert-extracted labels

Example paths¶

Role	Path
`base_image_dir`	`/path/to/MIMIC_CXR/physionet.org/files/mimic-cxr-jpg/2.0.0/files/`
`csv_path` (`mimic-cxr-2.0.0-metadata.csv`)	`/path/to/MIMIC_CXR/physionet.org/files/mimic-cxr-jpg/2.0.0/mimic-cxr-2.0.0-metadata.csv`
`label_csv_path` (`mimic-cxr-2.0.0-chexpert.csv`)	`/path/to/MIMIC_CXR/physionet.org/files/mimic-cxr-jpg/2.0.0/mimic-cxr-2.0.0-chexpert.csv`
`split_csv_path` (`mimic-cxr-2.0.0-split.csv`)	`/path/to/MIMIC_CXR/physionet.org/files/mimic-cxr-jpg/2.0.0/mimic-cxr-2.0.0-split.csv`

All CSVs live in the parent dir of files/ and are auto-discoverable from base_image_dir.