Skip to content

Latest commit

 

History

History
181 lines (149 loc) · 11.7 KB

dataset.md

File metadata and controls

181 lines (149 loc) · 11.7 KB
Raw

COVIDx CT Dataset

Quick Links

  1. Description
  2. Metadata
  3. Downloading the dataset
  4. Creating the dataset from scratch
  5. Data distribution
  6. Licenses and acknowledgements for the datasets used

Description

COVIDx CT, an open access benchmark dataset that we generated from open source datasets, currently comprises 201,103 CT slices from 4,501 patients. We will be adding to COVIDx CT over time to improve the dataset.

Labels for the images are obtained in one of three ways:

  1. Manual labelling or segmentation by radiologists (all validation and test images are labelled this way)
  2. Manual labelling by non-radiologists (these images are only included in the training set)
  3. Automatic labelling using a previous COVID-Net CT model (these images are only included in the training set)

metadata.csv indicates the labelling method used for each patient.

COVIDx CT is divided into "A" and "B" variants, the details of which are given below.

COVIDx CT-A

The "A" variant consists of cases with confirmed diagnoses (i.e., RT-PCR, radiologist-confirmed, etc.). COVIDx CT-3A comprises 425,024 CT slices from 5,312 patients.

COVIDx CT-B

The "B" variant contains all of the "A" variant and adds some cases which are assumed to be correctly diagnosed but could not be verified. COVIDx CT-2B comprises comprises 431,205 CT slices from 6,068 patients. Notably, the additional images included in this variant are only added to the training set, and as such the validation and testing sets are identical to those of the "A" variant.

Metadata

Metadata for each patient is included in metadata.csv. The metadata includes:

  • Patient ID
  • Data source
  • Country (if available)
  • Age & sex (if available)
  • Finding (Normal, Pneumonia, or COVID-19)
  • Verified finding, which indicates whether the finding is confirmed (Yes or No)
  • Slice selection, which indicates how slice selection was performed (either Expert, Non-expert, or Automatic)
  • View and modality (all are axial CT)

Some basic patient information from the dataset:

  • Countries: China, France, Russia, Iran, USA, Australia, Algeria, Italy, Scotland, Peru, Lebanon, England, Turkey, Belgium, Azerbaijan, Afghanistan, Ukraine
  • Age range (for cases which have age information): 0-94
  • Sexes: 3091 Unknown (50.9%), 1639 Male (27.0%), 1338 Female (22.1%)

Downloading the Dataset

The easiest way to use the COVIDx CT dataset is by downloading it directly from Kaggle. Links to different versions are provided below:

Note that COVIDx CT B variants are not available on Kaggle, and must be generated from scratch.

Creating the Dataset from Scratch

If you wish to construct the dataset from scratch, instructions for doing so are provided below.

Step 1: Downloading the Raw Data

We construct the "A" variant of the COVIDx CT dataset from the following publicly available data sources:

The following additional data source is leveraged to construct the "B" variant of the dataset:

To build COVIDx CT, the datasets above must first be downloaded. Additional instructions for some of the data sources are listed below.

CNCB Data

The CNCB data may be downloaded here. All CP, NCP, and Normal zip files should be downloaded, as well as unzip_filenames.csv and lesion_slices.csv. Once downloaded and extracted, the data should have the following structure:

  • <CNCB root>/
    • CP/
    • NCP/
    • Normal/
    • unzip_filenames.csv
    • lesions_slices.csv

LIDC-IDRI Data

The LIDC-IDRI challenge data may be downloaded here. Once downloaded, a .pylidcrc file must be created so that the pylidc package can locate the data. The .pylidcrc file must be located in the user's home directory, and has the following format:

[dicom]
path = /path/to/LIDC-IDRI
warn = True

iCTCF Data

The iCTCF data may be downloaded here. All patient cases listed in ictcf_metadata.csv should be downloaded in JPEG form. Once downloaded and extracted, the data should have the following structure:

  • <iCTCF root>/
    • Patient 1/
      • CT/
        • IMG-0001-00001.jpg
        • ...
    • Patient 2/
      • CT/
        • IMG-0001-00001.jpg
        • ...
    • ...

Step 2: Setting up the Notebook

Before constructing the dataset, additional packages must be installed:

The dataset is constructed from the downloaded sources using create_COVIDx_CT.ipynb. Several variables in this notebook must be set to reflect the locations of the raw data sources:

  • CNCB_EXCLUDE_FILE: this points to the file cncb_exclude_list.txt, and should not be modified unless this file is moved from its default location.
  • CNCB_EXTRA_LESION_FILE this points to the file cncb_extra_lesions_slices.csv, and should not be modified unless this file is moved from its default location.
  • CNCB_DIR: this should be set to the location of the prepared CNCB dataset, with Normal, CP, and NCP directories as well as metadata files.
  • RADIOPAEDIA_CORONACASES_CT_DIR: this should be set to the location of the COVID-19-CT-Seg_20cases directory from the COVID-19 CT Lung and Infection Segmentation Dataset.
  • RADIOPAEDIA_CORONACASES_SEG_DIR: this should be set to the location of the Infection_Mask directory from the COVID-19 CT Lung and Infection Segmentation Dataset.
  • RADIOPAEDIA_META_CSV: this points to the file radiopaedia_metadata.csv, and should not be modified unless this file is moved from its default location.
  • RADIOPAEDIA_EXCLUDE_FILE: this points to the file radiopaedia_exclude_list.txt, and should not be modified unless this file is moved from its default location.
  • LIDC_META_CSV: this points to the file lidc_idri_metadata.csv, and should not be modified unless this file is moved from its default location.
  • COVID_19_20_DIR: this should be set to the location of the Train directory from the COVID-19-20 challenge dataset.
  • TCIA_COVID_META_CSV: this points to the file tcia_covid_metadata.csv, and should not be modified unless this file is moved from its default location.
  • TCIA_DIR: this should be set to the location of the CT_Images_in_COVID-19_August_2020 directory from the NIH TCIA dataset.
  • COVID_CTSET_META_CSV: this should be set to the location of the file Labels&Detailes/Patient_details.csv from COVID-CTSet.
  • COVID_CTSET_DIR: this should be set to the location of the Train&Validation directory from COVID-CTSet.
  • ICTCF_META_CSV: this points to the file ictcf_metadata.csv, and should not be modified unless this file is moved from its default location.
  • ICTCF_DIR: this should be set to the location of the directory containing iCTCF patient cases in JPEG form (i.e., case directories for Patient 1, Patient 2, etc.).
  • COVID_CT_MD_DIR: this should be set to the root directory containing the {COVID-19, Cap, Normal} Cases directories from COVID-CT-MD.
  • COVID_CT_MD_INDEX_CSV: this should point to COVID-CT-MD's Index.csv file.
  • COVID_CT_MD_META_CSV: this points to the file covid_ct_md_metadata.csv, and should not be modified unless this file is moved from its default location.
  • COVID_CT_MD_LABEL_CSV: this should be set to the location of the file Slice-level-labels.npy from COVID-CT-MD.
  • STOIC_META_CSV: this points to the file stoic_metadata.csv, and should not be modified unless this file is moved from its default location.
  • STOIC_DIR: this should be set to the location of the data/mha directory of the STOIC data.
  • STONYBROOK_META_CSV: this points to the file stonybrook_metadata.csv, and should not be modified unless this file is moved from its default location.
  • STONYBROOK_DIR: this should be set to the location of the COVID-19-NY-SBU directory of the Stony Brook data (the inner directory containing patient cases).
  • OUTPUT_DIR: this should be set to the directory in which the final dataset will be created.

To create the experimental "B" variant of the dataset, additional paths must be set:

  • MOSMED_CT_DIR: this should be set to the location of the COVID19_1110/studies directory from MosMedData.
  • MOSMED_SEG_DIR: this should be set to the location of the COVID19_1110/masks directory from MosMedData.
  • MOSMED_META_CSV: this points to the file mosmed_metadata.csv, and should not be modified unless this file is moved from its default location.

Step 3: Running the Notebook

Once the notebook is prepared, the dataset is constructed by simply running all cells in the notebook. Notably, create_COVIDx_CT.ipynb will not re-create files which were previously created, allowing for the dataset construction to be interrupted without having to recreate all the files.

After running the construction cells, there is an optional check at the end of the notebook to ensure that all files were created successfully.

Data Distribution

COVIDx CT-3A

Chest CT image distribution

Type Normal Pneumonia COVID-19 Total
train 35996 26970 294552 357518
val 17570 8008 8147 33725
test 17922 7965 7894 33781

Patient distribution

Type Normal Pneumonia COVID-19 Total
train 321 592 3336 4249
val 164 202 194 560
test 164 138 201 503

COVIDx CT-3B

Chest CT image distribution

Type Normal Pneumonia COVID-19 Total
train 35996 26970 300733 363699
val 17570 8008 8147 33725
test 17922 7965 7894 33781

Patient distribution

Type Normal Pneumonia COVID-19 Total
train 321 592 4092 5005
val 164 202 194 560
test 164 138 201 503