Training and test data sets

Post here all questions about the challenge data sets.

Both input and label stacks can be visualized using several publicly available tools. For instance:

In Image and /Fiji you can change the lookup table of the label images to facilitate their visualization under Image > Lookup Tables.

You can also open them in Matlab as multipage TIFFs:

Please, feel free to post other available tools.

Alessandro Guisti and Dan Ciresan, the winners of the 2012 challenge on 2D segmentation of EM images, have kindly provided the 2D probability maps of this year's data sets based on in/out classification (membrane and non-membrane pixels).

This is a completely optional data that you can use as part of your segmentation method. 

Each slice was computed as the average of 8 images, i.e. the outputs of two different nets applied to 4 different rotations of the input
slice. No post-processing was applied.

The training ground truth data was preprocessed this way:

  • set up initial membrane labels as the union of:​
    •  all pixels with label 0
    • all pixels at the edge of any nonzero object.
  • shave off at most 10 pixels from all membranes (this thins most membranes to 1-pixel width, but does not destroy large areas with label 0).
  • dilate the resulting image using a 2-pixel radius disk kernel, which results in most membranes now having a width of about 4-5 pixels, which roughly matches the observed real membrane thickness.

The membrane probabilities were computed using deep convolutional networks as described in the following paper presented in NIPS 2012:

If you decide to use these probabilities in your segmentation methods, please include the next citation (BibTex format) on your work:

  title={Deep neural networks segment neuronal membranes in electron microscopy images},
  author={Dan Ciresan and Alessandro Giusti and Luca M. Gambardella and Juergen Schmidhuber},
  booktitle={Proceedings of Neural Information Processing Systems},