---
dataset_info:
  features:
    - name: image
      dtype: image
    - name: labels
      dtype: string
    - name: filename
      dtype: string
  splits:
    - name: train
      num_examples: 77967
    - name: validation
      num_examples: 8557
    - name: test
      num_examples: 25596
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
      - split: validation
        path: data/validation-*.parquet
      - split: test
        path: data/test-*.parquet
task_categories:
  - image-classification
task_ids:
  - multi-label-image-classification
language:
  - en
tags:
  - medical
  - chest-xray
  - radiology
  - deep-learning
  - big-data
  - parquet
  - data-only
license: unknown
size_categories:
  - 100K<n<1M
pretty_name: NIH Chest X-ray14 (320x320, Processed for CheXVision)
---

# NIH Chest X-ray14 - 320x320 Processed for CheXVision

## Project Resources

- [GitHub repository](https://github.com/arudaev/chexvision)
- [Presentation deck](https://arudaev.github.io/chexvision/)
- [Live demo](https://huggingface.co/spaces/arudaev/chexvision-demo)
- [Scratch model](https://huggingface.co/arudaev/chexvision-scratch)
- [DenseNet model](https://huggingface.co/arudaev/chexvision-densenet)

This dataset repackages the raw NIH Chest X-ray14 source dataset from
[alkzar90/NIH-Chest-X-ray-dataset](https://huggingface.co/datasets/alkzar90/NIH-Chest-X-ray-dataset)
into a data-only Parquet dataset for the CheXVision project.

## Dataset Summary

- Source format: 12 ZIP archives of original chest X-ray images plus CSV manifests
- Output format: data-only Parquet shards under `data/`
- Resolution: 320x320 RGB
- Columns: `image`, `labels`, `filename`
- Split contract: `train`, `validation`, `test`
- Approximate local Parquet size produced in this run: `7.4 GB`

## Splits

| Split | Images |
|-------|-------:|
| Train | 77,967 |
| Validation | 8,557 |
| Test | 25,596 |

## Schema

- `image`: 320x320 RGB image payload
- `labels`: pipe-delimited pathology labels, or `No Finding`
- `filename`: original NIH image filename

## Processing Notes

- Source split manifests come from `train_val_list.txt` and `test_list.txt`
- Validation membership uses the same stable hash-bucket logic as the live
  `HlexNC/chest-xray-14` dataset
- Hidden `__MACOSX` ZIP entries and non-image members are ignored
- Truncated-but-readable images are kept; truly unreadable files are skipped
- This repo intentionally ships no `load_dataset.py` script so it remains a
  data-only dataset that works with the modern HF dataset viewer

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("HlexNC/chest-xray-14-320")
print(dataset)
```

## Provenance

Built by the Kaggle kernel `hlexnc/chexvision-resize-320` for the
[CheXVision](https://github.com/arudaev/chexvision) project.