RSNA Pneumonia Processed Dataset

The RSNA Pneumonia Detection Dataset has been carefully pre-processed to streamline machine learning workflows, enabling faster and more efficient model training. Originally provided in DICOM format, the images have been converted into PNG format for compatibility and ease of use. The dataset preserves essential bounding box annotations by converting them into binary mask images, ensuring seamless integration into various training pipelines. Additionally, metadata, including patient details and bounding box coordinates, has been meticulously organized into CSV files, simplifying dataset exploration and feature engineering.

Folder Structure for Easy Navigation

The dataset is organized into a structured folder system, designed for straightforward usage:

1. Train

• Images: Pre-processed PNG images for training.
• Masks: Binary mask images corresponding to the bounding boxes in the training set.

2. Test

• Contains pre-processed PNG images designated for testing.

3. Metadata Files

• Train_metadata.csv: Includes patient IDs, bounding box coordinates, and pneumonia presence labels for the training set.
• Test_metadata.csv: Contains patient IDs and relevant details for the test set.

Key Data Fields

The dataset provides detailed annotations and metadata to support robust model development:

• patientId: Unique identifier for each patient or image.
• x, y, width, height: Bounding box dimensions specifying regions of interest.
• Target: Binary label indicating pneumonia presence (1 = Pneumonia, 0 = No Pneumonia).
• class: Categorization of conditions:
  • Lung Opacity
  • Normal
  • No Lung Opacity / Not Normal
• age: Patient’s age at the time of imaging.
• sex: Patient’s gender (e.g., Male, Female).
• modality: Imaging method (e.g., X-ray, CT Scan).
• position: Patient’s position during imaging (e.g., AP for Anterior-Posterior, PA for Posterior-Anterior).

Dataset Distribution

The dataset includes a diverse range of cases, distributed as follows:

• No Lung Opacity / Not Normal: 11,821 images
• Normal: 8,851 images
• Lung Opacity: 9,555 images