Training Data Release

#1

Happy belated world kidney day! The training data has been released on GitHub. Some things to note:

  • We opted for the NIFTI file format over the previously planned DICOM due to its greater simplicity, faster read times, and better support of compression. Nibabel is a great package for reading and manipulating this filetype in Python.
  • We’re making use of Git Large File Storage (git-lfs). Make sure you initialize Git LFS on your system before cloning!
  • You all have until April 5 to discover and report issues with the segmentation labels. After that, we will attempt to address all the concerns by April 15, and then the data will be frozen until the MICCAI 2019 deadline to provide stability.
  • Since this data was collected during routine clinical practice from many centers, the voxel spacing is quite variable. If you would be interested in a version of the data transformed and interpolated to a constant spacing for all patients, please let us know here.
  • For a more comprehensive description of this data and how it was annotated, we have written a manuscript data descriptor. It will appear on arxiv.org early next week, and we will provide a link to it once it does.
2 Likes
#2

The amended training data has been released! A pull or clone will now bring down the most up to date version of the data. You can find a changelog here.

We had hoped to release the fixed-spaced data today as well but unfortunately we’re still working out some kinks in the interpolation method. Our new target is next Monday, April 22. Apologies for the delay!

1 Like
#3

Hi! Thank you for new data release

Can you please explain how to interpretate the affine matrix in this dataset?

1 Like
#4

In short, multiplication by the affine matrix should take an homogeneous coordinate in voxel space, i.e. (i,j,k,1), and transform it to a coordinate in the real world of the form (x,y,z,1). This article has more information.

In the original training data release, I made the mistake of thinking that the convention was to transform to (z,y,x,1). In the amended release, I’ve just realized that I made the mistake of erroneously leaving ones on the first and third diagonal elements. I’m really sorry about this, and it has been fixed in the latest commit.

1 Like