metadata

configs:
  - config_name: main_data
    data_files: DL_info.csv
license: cc-by-sa-4.0

NIH DeepLesion Dataset

Introduction

The DeepLesion dataset contains 32,120 axial computed tomography (CT) slices from 10,594 CT scans (studies) of 4,427 unique patients. There are 1–3 lesions in each image with accompanying bounding boxes and size measurements, adding up to 32,735 lesions altogether. The lesion annotations were mined from NIH’s picture archiving and communication system (PACS).

Git Repository Structure

Git Branches:

Images_png: This branch contains PNG image files named in the format {patient index}_{study index}_{series index}_{slice index}.png. Sub-folders are indicated by using / or \ after the last underscore. All images are saved as unsigned 16-bit integers. To revert pixel intensity to its original Hounsfield Unit (HU) values, subtract 32768 from the pixel intensity (see scripts/apply_window.py for example). 3D context is also included, featuring an extra 30mm of slices both above and below the key slice for a thorough examination.
Images_png_wn: All slices have been converted to 8-bit PNG format, intensity-windowed according to DL_info.csv, and rescaled to the range [0-255] for use in machine learning applications.
Images_nifti: 3D NIfTI sub-volumes in .nii.gz format, derived from provided 2D 16-bit PNG images.
Images_zips: 56 compact zip files available for download. Utilize batch_download_zips.py script to facilitate batch downloads of these files.
Key_slices: Key slices overlaid with lesion annotations for review purposes.

Main Branch Contents:

DL_info.csv: DeepLesion annotations and metadata.
scripts/apply_window.py: Python code to apply intensity-windowing and normalization to the 16-bit PNG images.
scripts/batch_download_zips.py: Python code to batch downloads Images_png via 56 compressed zip files.
scripts/DL_save_nifti.py: Python code for converting 2D 16-bit slices to 3D nifti volumes.

Citation

@article{yan2018deeplesion,
  title={DeepLesion: automated mining of large-scale lesion annotations and universal lesion detection with deep learning},
  author={Yan, Ke and Wang, Xiaosong and Lu, Le and Summers, Ronald M},
  journal={Journal of medical imaging},
  volume={5},
  number={3},
  pages={036501--036501},
  year={2018},
  publisher={Society of Photo-Optical Instrumentation Engineers}
}