DL-Spectral Challenge data and information

This record contains the materials for the DL-sparse-view CT Challenge

----------------------------------------------------------------------------------------
CONTENTS of data/
----------------------------------------------------------------------------------------

The training data for developing the neural networks is in the subfolder "data".
All files are compressed with gzip in order to facilate faster downloads.
Data are partitioned into four batches, which also facilates downloading of the
individual files. Data are in python numpy's .npy format.
After uncompressing with gunzip the .npy files can be read into python
with the numpy.load command, yielding single precision floating point arrays
of the proper dimensions.

In the "data" folder are:
Phantom_batch?.npy
These arrays are 1000x512x512.
1000 images of pixel dimensions 512x512.
These are the true images.

FBP128_batch?.npy
These arrays are 1000x512x512.
1000 images of pixel dimensions 512x512.
These are the FBP reconstructed images from the 128-view sinograms.

Sinogram_batch?.npy
These arrays are 1000x128x1024.
1000 sinograms of 128 projections over 360 degree scanning onto a 1024-pixel linear detector. 

There are four batches. Thus 4000 sets of data/image pairs are available for training
the neural networks for image reconstruction.
The goal is to train a network that accepts the FBP128 image (and/or the 128-view sinogram)
to yield an image that is as close as possible to the corresponding Phantom image.

----------------------------------------------------------------------------------------
CONTENTS of validation-data/
----------------------------------------------------------------------------------------
Data is in the same arrangement as in the "data/" folder except that there are only 10 cases.
As a result they data are not split into batches and they are not compressed.
Phantom_validation.npy
These arrays are 10x512x512.
10 images of pixel dimensions 512x512.
These are the true images for the validation stage.  !!!! KEEP THIS SECRET !!!!

FBP128_validation.npy
These arrays are 10x512x512.
10 images of pixel dimensions 512x512.
These are the FBP reconstructed images from the 128-view sinograms.

Sinogram_validation.npy
These arrays are 10x128x1024.
10 sinograms of 128 projections over 360 degree scanning onto a 1024-pixel linear detector. 

----------------------------------------------------------------------------------------
Contents of this folder
----------------------------------------------------------------------------------------
"smallbatch" data
metrics.py
converMatlab.py
README (you're reading this now)

A "smallbatch" set of data is in this folder, containing only 10 phantoms, fbp images, and sinograms.

These data files are for viewing and are used to demonstrate the
metrics that will be used to evaluate the submitted images for this Grand Challenge.
Running the program metrics.py will compare the FBP128 images against the ground truth (Phantom images).
Hopefully your network will yield images that have lower RMSEs!
The two metrics are mean image RMSE, and worst-case ROI RMSE for a 25x25 pixel ROI.
The formulas for these metrics are in [put appropriate url link here],
and the metrics.py code can also be inspected to see how the calculation is performed.

The contest data are in numpy's .npy format and test image submission should also use
this format. For matlab users, a script "convertMatlab.py" is included that shows how
to convert the "smallbatch" data to matlab's .mat format. Also, converting back to .npy
is shown in this script.