ml-workgroup / covid-19-image-repository

Anonymized dataset of COVID-19 cases with a focus on radiological imaging. This includes images (x-ray / ct) with extensive metadata, such as admission-, ICU-, laboratory-, and patient master-data.

COVID-19 Image Repository

This project aims to create an anonymized data set of COVID-19 cases with a focus on radiological imaging. This includes images with extensive metadata, such as admission-, ICU-, laboratory-, and patient master-data.

This repository contains image data from the Institute for Diagnostic and Interventional Radiology, Hannover Medical School, Hannover, Germany.

Feature Set

id	label	unit	comment / reference interval (precision)
1	patient_id		randomly generated patient id
2	image_id		randomly generated image id (filename)
3	sex	m/w
4	age	years	currently redacted (see below)
5	size	cm	currently redacted (see below)
6	weight	kg	currently redacted (see below)
7	admission offset	days	days since admission (begin of symptoms)
8	icu admission offset	days
9	death offset	days	days until death or null
10	modality		currently all images are chest radiographs
11	projection		ap, pa ...
12	lactate dehydrogenase	U/l	< 248 (5)
13	c-reactive protein	mg/l	<= 5 (1)
14	d-dimer	mg/l	0 - 0.5 (0.1)
15	coagulation factor XIII	%	70 - 140 (5)
16	neutrophils	Tsd/µl	1.5 - 7.7 (0.1)
17	lymphocytes	Tsd/µl	1.1 - 4.0 (0.1)
18	pO2	mmHg	(5)
19	pCO2	mmHg	(5)
20	corona test type

Age, size, and weight are currently redacted. We will publish this data when there are enough patients, that meaningful intervals can be chosen according to the concept of k-anonymity and l-diversity.

Offsets, i.e. admission offset and icu admission offset, are given as relative times in regard to the exam. Please consult the feature set table for units. I.e. an admission offset == -4 and icu admission offset == 6 would encode, that the patient was admitted to the hospital four days ago and was transferred to the ICU six days after the image was taken. Especially the admission offset can be noisy; please see FAQ #1.

All lab values (12 - 19) are given in intervals to protect patient identity. A value below the detection limit is denoted by -inf, above the detection limit by inf.

Download

We provide the raw, unprocessed, gray value image data as Nifti files. This is done to protect patient identity, as Dicom files are hard to anonymize. However, the files are too large to host on Github.

Additionally, we included downscaled versions of the Nifti images in the png folder.

Space and bandwith are kindly provided by the Open Telekom Cloud (Deutschen Telekom AG).

Download Links: https://doi.org/10.25835/0090041

• Version 2.0

• Version 1.0

  Do not use this version (see erratum 1). The data is provided for reproducibility reasons.

Observations (FAQ)

1. admission offset > icu admission offset
   Some of the cases suggest that the admission to the clinic was after admission to the ICU. This is an artifact due to using the specific (COVID) case data for determining the admission date (and therefore offset). The patient may have already been at the ICU before being diagnosed with COVID.

Errata

1. In Version 1.0 there is an issue with the date calculation (issue #6). This might lead to incorrect date offset calculations. We strongly advise to upgrade to a newer version.

License and Attribution

Master-, image- and laboratory-data of this repository are licensed under the Creative Commons Attribution 3.0 Unported (CC BY 3.0).

If you use this data, you must attribute the authors in any publication (DOI: 10.6084/m9.figshare.12275009). You may include the specific release or commit hash for reproducibility.

Contact Information

Please use the ticketing system where applicable. Otherwise, please use the following EMail address: [email protected]

covid-19-image-repository's Issues

How to determine whether it is negative or positive?

I would like to ask if it is possible to determine whether a patient's photo data is negative or positive through the data.csv file?

Assume all images are COVID positive?

If the patient is not admitted yet (positive admission offset) should we assume they are COVID positive? Possibly just not testing positive at the moment?

Lymphocyte_offset & neutrophil_offset meaning?

Could you provide more explanation about these two variables?

Some images cropped when adding to covid-chestray-dataset

Hi! While adding these images to @ieee8023's covid-chestxray-dataset, I cropped a few to cut out part of the black margin. I just wanted to share the filenames in case you were interested in doing the same thing.

00d96e05.png
17ad0a56.png
436a6348.png
4c0fcf57.png
4fed5061.png
6f7008af.png
76093afc.png
81089cb4.png
865336ed.png
a3111116.png
a4318ac9.png
b10c49ca.png
bace1e45.png
bb4c4038.png
cb706009.png

Thank you so much for making this resource available!

CR vs DX

For modality what is the difference here?

What does ldh_offset or d-dimer_offset mean?

I need to clarify what does the offset for these clinical findings mean? Is it the days before the exam or something else?

Radiological findings

While I appreciate your great work providing images in a public repo, not having labels for images is a problem in my point of view. I have read your guideline for assuming COVID-positive/negative based on admission-offset but some of the CXRs are not clearly having radiological findings of pneumonia.

For example, CXR with patient_id=d3fb252e and image_id=88859dc1 has an admission offset of 0 with most of the other metrics not reported; while it is assumed COVID-19 positive, I think CXR does not show specific findings related to pneumonia. If a radiologist's interpretation could be provided for the images it would be great.

date offset calculation

There is an internal issue with the date offset calculation. We are working on a fix and will update shortly.

admission_offset

I don't understand this field. It is the days since symptoms/admission?

For example patient 09ec853d with an admission_offset of -21 and an icu_admission_offset -20.0 ? So from this image they went to the ICU in 20 days and then were admitted to the hospital a day after?

Is there any update to this?

Were any more X-Ray images released? Or will any be released soon?

Adding these images to the covid-chestxray-dataset ?

Hey! This data release is awesome! Thanks for releasing all this data with the detailed metadata!

I have been building this dataset: https://github.com/ieee8023/covid-chestxray-dataset

I would like to merge your images into this dataset. Is this ok with you? Let me know if not. I don't want to steal the thunder of your dataset!

all the X-ray images in the png folder are of covid positive patients right ?