adlito/covid19-cxr-bm3d-denoising

Dataset Card for COVID-19 CXR BM3D Denoising

Dataset Summary

This dataset contains paired chest X-ray (CXR) images for supervised image denoising. It is built from the *COVID-19 Radiography Database* and, for each original CXR, provides:

• —the original (clean) image,
• —a synthetically noisy version,
• —a BM3D-denoised version (pseudo–ground truth).

The goal is to train and evaluate models that learn to reproduce BM3D from noisy CXRs.

Important: This is a derived dataset. The original images and their copyrights belong to the authors of the COVID-19 Radiography Database. Usage is restricted to academic / non-commercial purposes (see Licensing).

Source Data

Original Dataset

This dataset is derived from the COVID-19 Radiography Database (COVID-19 Chest X-ray images and Lung masks Database), available on Kaggle.

Original dataset characteristics (at the time of extraction):

• —3,616 COVID-19 positive CXRs
• —10,192 Normal CXRs
• —6,012 Lung Opacity CXRs (non-COVID lung infection)
• —1,345 Viral Pneumonia CXRs
• —PNG images, 299×299 pixels

Original authors: Chowdhury, Rahman, Khandakar, Mazhar, Kadir, Mahbub, Islam, Khan, Iqbal, Al-Emadi, Reaz, Islam, and collaborators.

Data Collection and Processing

Noise Model

Each original CXR is transformed into a noisy version using additive white Gaussian noise, with the standard deviation σ sampled uniformly between 0.01 and 0.1 in normalized [0,1] grayscale intensity space.

BM3D Denoising

For each noisy image, a BM3D-denoised image is computed using bm3d package.

The BM3D output serves as pseudo–ground truth for learning-based denoising methods.

Dataset Structure

A typical sample contains:

• —original: original CXR (from COVID-19 Radiography Database)
• —noisy: synthetically noised version of the CXR
• —bm3d: BM3D-denoised version of the noisy image

The noisy images are named using the pattern CLASS-IMAGEID-NOISELEVEL.png, where CLASS is the original label (for example, COVID or Normal), IMAGEID is the index of the corresponding clean image, and NOISELEVEL is the noise standard deviation σ multiplied by 1000 and rounded to an integer (for example, COVID-000123-050.png for σ = 0.05).