PET segmentation challenge using a data management and processing infrastructure

Data and format

Following recommendations by the TG211, the evaluation dataset will consist of a mixture of numerical simulations with realistic uptake distributions and a variety of tumor shapes, physical phantoms acquisitions, and real clinical images. Simulated images consist of simple synthetic images obtained by adding noise and blur to the ground-truth, as well as more realistic objects generated for example with GATE to better account for the physics of PET acquisition. Voxel-by-voxel ground-truth is available from the simulation. The physical phantoms will be zeolites with different sizes and shapes incorporated in an anthropomorphic phantom, with repeated acquisitions of the same objects. The surrogate of truth for these images is the corresponding high resolution CT thresholded as to obtain the exact known volume of the zeolite. Clinical images are lung and head and neck tumors, with either histopathology-based contours or a consensus of several manual delineations by experts to serve as surrogate of truth.

A few test images will be provided to challengers beforehand for testing and optimization purposes. The images used for the challenge itself will not be shared with challengers. They will have to integrate their algorithm in a container so it can be run directly on a platform (see pipeline integration section).


Data hosting and access

Challengers will access the dedicated web portal home page provided by the FLI-IAM infrastructure. On this portal, the challengers will get access to the cloud data hosting solution (Shanoir[2]) and will be able to:

  • Find information about the challenge
  • Register to the portal and for the challenge using single-sign-on services
  • Download the training data with ground truth directly
    from the portal challenge pages in .zip format (all raw data will be provided in the NIFTI format)


Training data

Several examples representative of the overall challenge dataset.

  1. Simulated images: 1 case simulated with GATE and 4 cases simulated with the SIMSET.
  2. Real phantom acquisitions: 3 different zeolites within an anthropomorphic phantom with 3 repeated acquisitions for each zeolite.
  3. Patient clinical data: 1 lung cancer case with manual delineation (statistical consensus of 3 experts), 2 cases with associated histopathology-derived contour: 1 lung cancer case reconstructed with two different voxel sizes, and 1 head and neck cancer case.