exalsius/NIH-Chest-XRay-Federated

NIH Chest X-ray Federated Learning Dataset

Federated learning splits designed for the [Cold Start:] Distributed AI Hack Berlin 2025.

The dataset is based on the NIH Chest X-ray14 dataset, which contains ~112,000 X-ray images from 30,805 unique patients, and models a federated learning scenario with non-IID characteristics across three hospitals, plus an out-of-distribution test set.

Dataset Description

The data was partitioned using a scoring algorithm that creates non-IID distributions:

1. Patient-level splitting: Each patient appears in only one hospital/split
2. Demographic biasing: Age and sex distributions vary across hospitals
3. Equipment simulation: AP/PA view ratios differ by hospital type
4. Pathology concentration: Each hospital has characteristic disease patterns
5. Train/eval/test split: 80/10/10 split within each hospital (patient-disjoint)

See the preparation script for implementation details.

Data Distribution

We partitioned the chest X-rays into hospital silos that reflect real-world data heterogeneity:

• Hospital A (Portable Inpatient): 42,093 train, 5,490 eval

  • Demographics: Elderly males (age 60+)
  • Equipment: AP (anterior-posterior) view dominant
  • Common findings: Fluid-related conditions (Effusion, Edema, Atelectasis)

• Hospital B (Outpatient Clinic): 21,753 train, 2,860 eval

  • Demographics: Younger females (age 20-65)
  • Equipment: PA (posterior-anterior) view dominant
  • Common findings: Nodules, masses, pneumothorax

• Hospital C (Mixed with Rare Conditions): 20,594 train, 2,730 eval

  • Demographics: Mixed age and sex
  • Equipment: PA view preferred
  • Common findings: Rare conditions (Hernia, Fibrosis, Emphysema)

Test Sets

The dataset includes 4 test sets:

• test_A: In-distribution test for Hospital A
• test_B: In-distribution test for Hospital B
• test_C: In-distribution test for Hospital C
• test_D: Out-of-distribution ICU/Critical Care data (age extremes, multi-morbidity)

All splits are patient-disjoint to prevent data leakage.

Usage

from datasets import load_dataset

# Load Hospital A data
hospital_a = load_dataset("exalsius/NIH-Chest-XRay-Federated", "hospital_a")
# Returns: DatasetDict({'train': Dataset, 'eval': Dataset})

# Load Hospital B
hospital_b = load_dataset("exalsius/NIH-Chest-XRay-Federated", "hospital_b")
# Returns: DatasetDict({'train': Dataset, 'eval': Dataset})

# Load Hospital C
hospital_c = load_dataset("exalsius/NIH-Chest-XRay-Federated", "hospital_c")
# Returns: DatasetDict({'train': Dataset, 'eval': Dataset})

# Load test sets
test_data = load_dataset("exalsius/NIH-Chest-XRay-Federated", "test")
# Returns: DatasetDict({'test_a': Dataset, 'test_b': Dataset, 'test_c': Dataset, 'test_d': Dataset})

Original NIH Dataset

@article{wang2017chestxray,
  title={ChestX-ray8: Hospital-scale Chest X-ray Database and Benchmarks on
         Weakly-Supervised Classification and Localization of Common Thorax Diseases},
  author={Wang, Xiaosong and Peng, Yifan and Lu, Le and Lu, Zhiyong and
          Bagheri, Mohammadhadi and Summers, Ronald M},
  journal={CVPR},
  year={2017}
}