TAIX-Ray¶

Modality: CXR (bedside AP, ICU cohort) | Format: 16-bit grayscale PNG (MONOCHROME2) | Dim: 2D | Labels: 8 findings, prospective structured expert grading (binary or ordinal)

Overview¶

TAIX-Ray is a German bedside intensive-care CXR dataset from University Hospital RWTH Aachen (Truhn Lab). It comprises 215,381 anteroposterior bedside radiographs from 47,724 ICU patients (median age 68) collected across 10 ICU wards on 18 Siemens Mobilett MIRA mobile radiography systems over 14 years (January 2010 – December 2023). During routine clinical reporting, 134 radiologists provided prospective, structured, itemized annotations for 8 clinically relevant findings (not NLP-extracted labels) using a standardized 5-point ordinal severity template (heart size uses 4 grades; all others use 5). The dataset supports two label modes: binary (severity ≥ 1 → 1) and ordinal (raw 0–4 grades, 0–3 for heart size). Two resolutions are distributed: 512-px longer-dim resized (bilinear) and original resolution. RadHarmony provides a separate dataset class for each resolution variant.

Citation: Truhn D, Geiger D, Siepmann R, von der Stück MS, Bressem KK, Kather JN, Kuhl C, Müller-Franzes G, Nebelung S. A comprehensive bedside chest radiography dataset with structured, itemized and graded radiologic reports. Scientific Data 2026;13:632. DOI 10.1038/s41597-026-07271-7. Repository: github.com/TruhnLab/TAIX-Ray.

Provenance note (corrected 2026-07-11): TLAIM is the HuggingFace organization name for "Truhn Lab AI Medicine" (a lab-name acronym, not a country code). Prior wiki text incorrectly described this as a Thai dataset; the source paper places all authors at RWTH Aachen (Germany) and Technical University of Munich, with Kather at TU Dresden. Images are 16-bit grayscale PNG stored as MONOCHROME2 with inversion applied where required; no VOI LUT windowing or intensity normalization is applied by the release.

Download¶

Available from HuggingFace: TLAIM/TAIX-Ray. Images are expected in a flat images/ directory with an annotation.csv.

Expected layout¶

TAIX-Ray/
  data_512/
    images/
      annotation.csv
      <image_id>.png
  data_original/
    images/
      annotation.csv
      <image_id>.png

Label columns¶

Column	Description
`atelectasis_left`	Left atelectasis
`atelectasis_right`	Right atelectasis
`heart_size`	Heart size abnormality
`pleural_effusion_left`	Left pleural effusion
`pleural_effusion_right`	Right pleural effusion
`pulmonary_congestion`	Pulmonary congestion
`pulmonary_opacities_left`	Left pulmonary opacities
`pulmonary_opacities_right`	Right pulmonary opacities

In binary mode: each column is 0 or 1. In ordinal mode: each column holds a severity grade (0 = absent, higher = more severe).

Constructor arguments¶

Both TAIXRay512Dataset (512px) and TAIXRayDataset (original resolution) share the same signature:

Argument	Type	Required	Default	Description
`base_image_dir`	`str`	Yes	`None`	`images/` directory — flat dir of `<UID>.png` files (e.g. `/data/TAIX-Ray/data_512/images/` for the 512-px variant, or `data_original/images/` for the original-resolution variant)
`csv_path`	`str`	No	auto	`annotation.csv`; auto-discovered
`label_mode`	`str`	No	`"binary"`	`"binary"` or `"ordinal"`

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`	Include `"mask"` in data dict
`output_report`	`bool`	No	`False`	Include `"report"` in data dict
`output_bbox`	`bool`	No	`False`	Include `"bbox"` and `"bbox_labels"` in data dict
`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¶

512px variant¶

from radharmony.dataset import TAIXRay512Dataset

ds = TAIXRay512Dataset(
    base_image_dir="/data/TAIX-Ray/data_512/images/",
    label_mode="binary",
    output_cls=True,
)
train_ds, val_ds = ds.get_datasets(n_splits=5)

Original resolution variant¶

from radharmony.dataset import TAIXRayDataset

ds = TAIXRayDataset(
    base_image_dir="/data/TAIX-Ray/data_original/images/",
    label_mode="ordinal",
    output_cls=True,
)

Harmonizer¶

from radharmony.harmonizer import TAIXRayHarmonizer

h = TAIXRayHarmonizer(
    csv_path="/data/TAIX-Ray/data_512/images/annotation.csv",
    base_image_dir="/data/TAIX-Ray/data_512/images/",
)
df = h.harmonize()
print(df.columns.tolist())
df.to_csv("taix_ray_harmonized.csv", index=False)

Load from saved harmonized CSV¶

import pandas as pd
from radharmony.dataset import TAIXRay512Dataset

ds = TAIXRay512Dataset(
    base_image_dir="/data/TAIX-Ray/data_512/images/",
    harmonized_df=pd.read_csv("taix_ray_harmonized.csv"),
    output_cls=True,
)

Harmonizer notes¶

label_mode="binary" binarises ordinal grades (0 = absent, >0 = present)
label_mode="ordinal" preserves the original severity grades
Both TAIXRay512Dataset and TAIXRayDataset use the same harmonizer and annotation format; they differ only in expected image directory

Outputs¶

Flag	Key	Shape	Notes
`output_cls=True`	`"cls"`	`(8,)`	Binary or ordinal labels depending on `label_mode`

Example paths¶

Role	Path
`base_image_dir` (512 px)	`/path/to/TAIX-Ray/data_512/images/`
`base_image_dir` (original)	`/path/to/TAIX-Ray/data_original/images/`
`csv_path` (512 px, `annotation.csv`)	`/path/to/TAIX-Ray/data_512/images/annotation.csv`
`csv_path` (original, `annotation.csv`)	`/path/to/TAIX-Ray/data_original/images/annotation.csv`