The data was kindly provided by medicalsegmentation.com. They are two radiologists from Oslo, who've done plenty of work scraping and segmenting CT images. Also, they have made a great product, where you can segment CT images in a browser. The magic of their tool is that it has several pretrained Unets that run in offline mode (no user data is going out) and make your work a hundred times faster. An additional contribution of lung masks was made by Johannes Hofmanninger. His approach on Github.

Dataset consists of two parts:

Medseg part

This is a dataset of 100 axial CT images from >40 patients with COVID-19 that were converted from openly accessible JPG images found HERE. The conversion process is described in detail in the following blogpost: Covid-19 radiology — data collection and preparation for Artificial Intelligence

images_medseg.npy - training images – 100 slices 512x512 size
masks_medseg.npy - training masks – 100 masks with 4 channels: (0 - "ground glass", 1 - "consolidations", 2 - "lungs other", 3 - "background" )
test_images_medseg.npy - test images – 10 slices 512x512 size

Radiopedia part

Segmented 9 axial volumetric CTs from Radiopaedia. This dataset includes whole volumes and includes, therefore, both positive and negative slices (373 out of the total of 829 slices have been evaluated by a radiologist as positive and segmented). These volumes are converted and normalized in a similar way as above.

images_radiopedia.npy - training images – 829 slices 512x512 size
masks_radiopedia.npy - training masks – 829 masks with 4 channels: (0 - "ground glass", 1 - "consolidations", 2 - "lungs other", 3 - "background" )

Predictions

Participants should predict 10 masks corresponding to test_images_medseg.npy with two classes: 0 - "ground glass", 1 - "consolidations".

Use of other datasets:

Participants can use any data available if they share it with others via posting a link in Discussion
For example, a good source of labeled COVID-19 CT data is available here: https://mosmed.ai