MouGam/nih-processed-dataset

NIH ChestX-ray14 — Preprocessed Dataset

Dataset Description

Preprocessed version of the NIH ChestX-ray14 dataset for multi-label thoracic disease classification.

Source

• Original Dataset: NIH ChestX-ray14
• Institution: NIH Clinical Center
• License: CC0 1.0 (Public Domain)

Citation

@inproceedings{wang2017chestx,
  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},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages={2097--2106},
  year={2017}
}

Preprocessing Pipeline

1. Quality Filtering

Images were filtered based on pixel intensity statistics from the original 112,120 images:

• Too dark (mean < 50): 61 images removed, 28 manually cropped (black border removal) and retained
• Too bright (mean > 195): 52 images removed, 2 manually cropped (white border removal) and retained
• Low contrast (std < 25 after CLAHE): 28 images removed, 1 retained (00004480_000.png)
• Total removed: 141 images
• Total retained: 111,979 images

2. Image Processing (applied to all retained images)

1. CLAHE (Contrast Limited Adaptive Histogram Equalization): clipLimit=2.0, tileGridSize=(8,8)
2. Resize: 224×224 pixels, Bilinear interpolation
3. Channel: Grayscale → 3-channel RGB (channel replication)
4. Format: PNG

3. Normalization

Not applied at save time. Apply ImageNet normalization at DataLoader time:

• mean = [0.485, 0.456, 0.406]
• std = [0.229, 0.224, 0.225]

4. Train/Test Split

• Method: Patient ID-based GroupShuffleSplit (85/15)
• Train: 96,359 images (26,152 patients)
• Test: 15,620 images (4,616 patients)
• Patient overlap: 0 (verified)

5. Cross-Validation

• Method: 5-Fold GroupKFold (Patient-wise) on train set
• Fold column included in train.csv (values 0–4)
• Fold-to-fold patient overlap: 0 (verified)

6. Multi-hot Encoding

14 diseases encoded in alphabetical order:

Index	Disease
0	Atelectasis
1	Cardiomegaly
2	Consolidation
3	Edema
4	Effusion
5	Emphysema
6	Fibrosis
7	Hernia
8	Infiltration
9	Mass
10	Nodule
11	Pleural_Thickening
12	Pneumonia
13	Pneumothorax

"No Finding" → all zeros [0,0,0,0,0,0,0,0,0,0,0,0,0,0]

Directory Structure

processed/
├── README.md
├── available/
│   ├── images/          # 111,979 preprocessed images (224×224×3 PNG)
│   ├── data.csv         # Full metadata + multi-hot encoding
│   ├── train.csv        # Train split (96,359 rows, includes fold column)
│   └── test.csv         # Test split (15,620 rows)
└── unavailable/
    ├── images/          # 141 filtered-out original images
    └── data.csv         # Metadata for filtered images

CSV Columns

Column	Description
Image Index	Filename (e.g., 00000001_000.png)
Finding Labels	Pipe-separated labels (e.g., Atelectasis|Effusion)
Follow-up #	Follow-up visit number
Patient ID	Unique patient identifier
Patient Age	Age in years
Patient Gender	M or F
View Position	PA or AP
OriginalImage[Width,Height]	Original image dimensions
OriginalImagePixelSpacing[x,y]	Pixel spacing
Atelectasis ... Pneumothorax	Multi-hot encoded disease columns (0 or 1)
fold	(train.csv only) Cross-validation fold index (0–4)