PadChest¶

Modality: CXR | Format: PNG | Dim: 2D | Labels: 193 (174 findings + 19 diagnoses) | Reports: Spanish

Overview¶

PadChest (BIMCV) comprises 160,868 chest X-rays (the distributed label CSV has 160,861 rows) from ~67,000 patients at Hospital San Juan de Alicante, Spain (2009–2017). Labels are NLP-derived from Spanish radiology reports plus manual review; the published label vocabulary contains 193 distinct labels — 174 radiographic findings plus 19 differential diagnoses (much richer than the 14-class CheXpert/ChestX-ray14 taxonomy). Each row also carries a pre-processed (stemmed/lemmatized) Spanish radiology report.

RadHarmony provides a single dataset class for this dataset: - PadChestDataset — full dataset (160,758 rows after dropping 103 rows with NaN Labels); classification + Spanish report.

Download¶

PadChest is hosted at the BIMCV Nextcloud share. The ?accept=zip whole-folder download is a dynamically streamed Zip64 that does not support resume — for a ~1 TB download this is fragile; prefer fetching the 52 per-shard zips individually (each ~20 GB, resumable via standard wget -c).

Authoritative primary source: BIMCV (Barcelona Supercomputing Center / Universitat Politècnica de València). Do not use third-party Kaggle mirrors — they are partial re-uploads.

The PNGs live in 52 sibling subdirectories under images/. This is the same structural-exception pattern as ChestX-ray14's images_001/.../images_012/: base_image_dir is the deepest stable directory (images/) and the relative image_path carries the subdir name (<ImageDir>/<ImageID>).

Expected layout¶

PadChest/
  PADCHEST_chest_x_ray_images_labels_160K_01.02.19.csv.gz
  LICENSE.md
  README.md
  Verify_Zips_ImageCounts.csv.xlsx
  images/
    0/   ~3000 PNGs
    1/   ~3000 PNGs
    ...
    50/  1 PNG       (legitimate — see notes)
    54/  4297 PNGs   (catch-all; slots 51/52/53 don't exist)

Label columns¶

193 labels (174 radiographic findings + 19 differential diagnoses), NLP-derived from Spanish reports. All exposed as binary one-hot columns. A representative subset:

Column	Description
`normal`	No finding (50,616 positives)
`copd_signs`	COPD signs (23,280)
`cardiomegaly`	Enlarged heart (15,022)
`unchanged`	Unchanged from prior study (14,349)
`aortic_elongation`	Aortic elongation (10,824)
`pleural_effusion`	Pleural effusion (10,026)
`scoliosis`	Scoliosis (8,335)
`pneumonia`	Pneumonia (8,037)
`interstitial_pattern`	Interstitial pattern (7,831)
`chronic_changes`	Chronic changes (7,343)
`infiltrates`	Infiltrates (7,156)
`consolidation`	Consolidation
`pneumothorax`	Pneumothorax
`atelectasis`	Atelectasis
`nodule`	Pulmonary nodule
`mass`	Pulmonary mass
`tuberculosis`	Tuberculosis
`pulmonary_fibrosis`	Pulmonary fibrosis
`nsg_tube`	NSG tube (device)
`endotracheal_tube`	Endotracheal tube (device)
`pacemaker`	Pacemaker (device)
… (172 more)

The full alphabetically-sorted list of 193 snake_cased columns is in configs/datasets/padchest.json and in PadChestDataset.LABEL_COLS. Counts above are post-row-drop totals (after the 103 NaN-Labels rows are removed).

Constructor arguments¶

Argument	Type	Required	Default	Description
`base_image_dir`	`str`	Yes	`None`	The `images/` directory containing `0/`, …, `50/`, `54/` sibling subdirs (e.g. `/data/PadChest/images/`)
`csv_path`	`str`	No	auto	`PADCHEST_chest_x_ray_images_labels_160K_01.02.19.csv.gz` (or `.csv`); auto-discovered
`drop_uncertain`	`bool`	No	`True`	Drop ~3,151 rows flagged with `exclude=1` or `suboptimal_study=1`. Set to `False` to train on the full noisy distribution.

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 and is ignored
`output_report`	`bool`	No	`False`	Include `"report"` (Spanish text) in data dict
`output_bbox`	`bool`	No	`False`	Not supported — warns and is ignored
`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 saved harmonizer pickle

Dataset constructor¶

import torch
from radharmony.dataset import PadChestDataset

ds = PadChestDataset(
    base_image_dir="/data/PadChest/images/",
    output_cls=True,
    output_report=True,
    dtype=torch.float32,
)
train_ds, val_ds = ds.get_datasets(n_splits=5)

Harmonizer¶

from radharmony.harmonizer import PadChestHarmonizer

h = PadChestHarmonizer(
    csv_path="/data/PadChest/PADCHEST_chest_x_ray_images_labels_160K_01.02.19.csv.gz",
    base_image_dir="/data/PadChest/images/",
)
df = h.harmonize()
print(df.columns.tolist())   # patient_id, study_id, image_path, view_position, report, + 193 labels
df.to_csv("padchest_harmonized.csv", index=False)

Load from saved harmonized CSV¶

import pandas as pd
from radharmony.dataset import PadChestDataset

ds = PadChestDataset(
    base_image_dir="/data/PadChest/images/",
    harmonized_df=pd.read_csv("padchest_harmonized.csv"),
    output_cls=True,
    output_report=True,
)

Harmonizer notes¶

Label parsing: the raw Labels column stores a Python literal string list per row (e.g. "['cardiomegaly', 'pleural effusion']"). The harmonizer parses with ast.literal_eval, strips per-label whitespace (some raw labels have leading-space variants like ' pneumonia' vs 'pneumonia' — both merge into the same one-hot column), and one-hot encodes against the 193-label vocabulary.
Dropped rows: 103 of 160,861 rows have NaN Labels and are always dropped. With the default drop_uncertain=True, an additional ~3,151 rows flagged by the annotators with exclude or suboptimal_study are removed, leaving 157,607 rows. Pass drop_uncertain=False on either the harmonizer's harmonize() call or the dataset constructor to keep them and harmonize 160,758 rows.
No per-label uncertainty: unlike CheXpert, PadChest has no -1 uncertain marker for individual findings. The Labels column is positive-only — a label is either in the list (encoded as 1) or absent (encoded as 0). Annotators expressed uncertainty at the image level via the exclude / suboptimal_study meta-labels documented above.
Reports: Spanish, pre-processed (stemmed/lemmatized) by the PadChest authors. Stored inline in the CSV's Report column — the harmonizer copies the text directly to report rather than reading separate report files. 17 rows have NaN reports (labels were available but no narrative text).
View position: uses the cleaned Projection column (PA / L / AP / AP_horizontal / COSTAL) rather than the raw ViewPosition_DICOM (POSTEROANTERIOR / LATERAL / ...). Projection is closer to what other CXR harmonizers expose and is fully populated.
Image paths: ImageDir is stored as an int in the CSV (0..50, 54); the harmonizer casts to str so the relative path is e.g. "0/<hash>_<slug>.png". Slot 50 has exactly one PNG and slot 54 is the catch-all that absorbs the missing 51/52/53 — both are legitimate quirks of the official distribution and verified against the per-zip *.unzip-l.txt listings.
Bboxes: PadChest's Localizations column contains anatomic region strings ('loc basal', 'loc cardiac', …), not pixel-coordinate bounding boxes. PadChestDataset therefore does not expose output_bbox.

Outputs¶

Flag	Key	Shape	Notes
`output_cls=True`	`"cls"`	`(193,)`	Multi-label binary
`output_report=True`	`"report"`	`str`	Spanish pre-processed text (NaN where the source report was missing)