General Information

Multi-dimensional Diffusion MRI (MUDI) Challenge will focus on combined diffusion-relaxometry MRI data analysis. The main goals of the challenge are to find the optimal data sub-sampling strategy to obtain MUDI data with high information content efficiently and to investigate suitable data representations to make use of the full parameter space.

The challenge data will consist on unique 3T whole-brain datasets (5 for training and 3 testing), with multiple TE, TI, diffusion weightings/encodings for a total acquisition time of 1 hour and 30 minutes (per subject) approximately.

For training data we will provide the fully-sampled complex image data (minimal pre-processing including denoising, motion and eddy current corrections will be applied to the images).

Data is provided by the Centre for the Developing Brain, Centre for Medical Engineering, King's College London, UK. All interested participants need to fill out the data sharing agreement (OA-Agreement) and the interest form that can be found here and the data will then be sent via file link.

The challenge results and discussion will be presented within the CDMRI'19 full-day workshop.

MUDI Challenge 2019 - Data

Training data:

The training data will include whole-brain MR datasets from 5 healthy adult volunteers. The datasets were acquired on a clinical 3T Philips Achieva Scanner using a 32-channel head coil. Resolution 2mm isotropic, Multiband 2, SENSE 2. In total, 3000 volumes (3 b x 7 IT x 4 TEs) will be provided in fully anonymized form (nifti format). A text file provided with the data will specify bvector, bvalue, inversion time and echo time for each volume in the same order.

Pre-processed data: The raw MR data will be processed with a minimal preprocessing pipeline, as follows. After denoising of the magnitude data using MP-PCA denoising implemented in MRtrix, motion/eddy current correction will be performed using FSL EDDY) and EPI distortion correction using FSL TOPUP. Scripts will be provided after the closure of the challenge.

Raw and pre-processed data will be provided for the training data. We note that evaluation of the challenge will be performed only based on the pre-processed data. However, the participants are welcome to run a customized data processing pipeline for method design and for Task 3.

Test data:

The test data will include whole-brain MR datasets from 3 additional healthy adult volunteers. The data was acquired and pre-processed in the same way as the training data (see above). Before the end of the challenge, only sub-sampled pre-processed datasets will be provided upon a request from the participants (see Challenge Task 0 for details). The fully-sampled raw and pre-processed data will be made available after the challenge.

MUDI Challenge 2019 - Tasks

Task 0: Select “the best 500 volumes” you like!

Choose a subsampling of the signal while respecting the constraint of 500 total volumes, corresponding to roughly 15 minutes and compatible with a clinical acquisition.

In this task, the participants will explore the training datasets and decide a sub-sampling strategy that can best predict the fully-sampled data.

How does it work:

The participants need to choose a subset of the fully-sampled data, with a maximum number of volumes=500. The participants can freely choose a combination of any b-value, directions, inversion times and echo times. They will do this by sending a text file (GroupName_MUDI_subset.txt) with the number of the volumes which are part of the chosen subset. The ordering starts at 0 and the number of the selected volumes needs to be comma-separated, an example of the file selecting 10 volumes is here. This task will not be evaluated, but it is fundamental to the next ones. Bear in mind: the participants will need to use this sub-sampled data to reconstruct the fully-sampled data as close as possible. (See Task 1 for the Challenge evaluation criteria.) We will only provide the corresponding subset of samples from the test data for you to perform the tasks.

Task 1: How much can you predict?

Provide us with the predicted signal on unseen (test) data using the subsampled data (N = 500 volumes) that we will provide you after completion of task 0. Now push it even more: subsample your subsampling even further to provide us with predicted signal based on 250, 100 and 50 volumes respectively. Note that we will evaluate the results you provide us with, e.g. if you provide only the 500 volumes case than we only evaluate that for the “500 volumes category”.

How does it work:

After you submit task 0, we will give you access to the unseen data related to only the subsamples that you have requested. Based on these, you will have to provide us with your reconstructed full-brain signals for the 500, 250, 100 and 50 volumes cases by deadline. You will also need to provide the IDs for the selected volumes for the 250,100 and 50 subsampling.

How do we evaluate that:

We will have the fully-sampled unseen signal available, so we can compare this with your predictions.

1. We will compute the mean squared error (MSE) for each prediction independently (500, 250, 100 and 50 volumes) and produce a ranking of the submissions [evaluation 1]. Moreover we will also sum the MSEs of the predictions together.

2. We will also evaluate the MSE within specific challenging ROIs, e.g. to assess robustness with respect to fiber-crossing, partial voluming, anatomical boundaries, etc [evaluation 2].

We might also assess the performance based on other metrics, e.g. the Cramer-Rao lower bound.

Task 2: How well can you segment?

Use your 500 volumes...

How does it work:

Provide us with 3 brain volume fraction maps, one for each main tissue (WM, GM, CSF) containing the relative quantity of the partial volume “occupied” by such tissue. For example, this could be based on a signal model (e.g. intra-axonal, extra-axonal, CSF compartments). Again, it could be based on a classification method. Submissions may make use of the fully-sampled training data to learn the mapping to the subsampled data. These are just examples, you can achieve the goal as you prefer.

How do we evaluate that:

This task will not be evaluated but discussed and voted during the workshop. There might be a nice reward [a bonus drink] for the best segmentation! [qualitatively assessed] ;)

Task 3: What else can you do?

Feel free to explore the data. Use the data (subsampled/post-processed or not) to produce scalar maps, tractography, new metrics, etc. You can even try to push the segmentation even further, by identifying other sub-tissues/structures. Propose new ways to look at the data and(or) maps where a combined acquisition has an impact.

MUDI Challenge 2019 - Other

Participation to all the tasks is not mandatory. Task 1 (evaluation 1 and evaluation 2) will be used to rank the winner of the Challenge but all contributions are welcomed.

Everyone who participates will be part of a potential paper and some abstracts.

Important dates

June 30th, 2019 First data release
July 11th, 2019 Full data release
September 15th, 2019 Challenge submission hard deadline
October 17th, 2019 Challenge and workshop in Shenzhen, China