White matter modelling challenge



This challenge is part of the ISBI Challenges, and it aims to identify the mathematical model for diffusion MRI that best describes the signal from in-vivo human brain white matter acquired on the Connectom scanner with gradients of up to 300mT/m.

The challenge looks at a simple situation, ROI voxels from the genu in the corpus callosum, where the fibres are approximately straight and parallel, and from the fornix, where the configuration of fibres is more complex. The picture below shows the location of these ROIs in the brain.

We provide a broad set of measurements that covers the set of b-values and diffusion times as widely as possible. Challenge participants have access to three-quarters of the whole dataset from each genu and fornix ROI. Though we will explore the learnt structures in both regions, the winning model, however, is the one which predicts just the remaining 'unseen' quarter of the genu data most closely.

The recent development of human MR systems with 300mT/m gradients, in particular the Connectom scanner, is a major step towards the long-term translation of microstructure imaging techniques to widespread clinical practice. Dyrby et al. illustrate the benefits of stronger gradients for mapping axon diameters and other parameters in WM using fixed post-mortem tissue and a small-bore animal imaging system. The first experiments verifying those findings on live human subjects are now beginning to emerge from the Connectom scanner [McNab et al., Huang et al.].


The challenge of finding an accurate model for the whole range of diffusion MRI measurements is important, because it helps us identify the range of tissue parameters we can potentially estimate. Ultimately, that will improve the diagnosis and monitoring of neurological conditions.

The challenge builds on previous studies that compare a range of compartment models using standard model selection criteria: Panagiotaki et al. (NeuroImage 2012), provided a taxonomy and compared 47 compartment models; Ferizi et al.(ISBI'13) performs a similar experiment using data from a live human subject. Ferizi et al.(MICCAI'13) explored a different class of models that aim to capture fibre dispersion.

A much wider variety of models are of course available, including other compartment models (e.g. Ball-and-Rackets from Sotiropoulos et al. and NODDI from Zhang et al.), other parametric models (e.g. the Gamma Diffusivities Model from Jbabdi et al., DIAMOND from Scherrer et al., the Kurtosis Model from Jensen et al.), and a variety of non-parametric models, e.g. Cumulant Expansion (see chapter 10 by Valerij Kiselev), SHOR and other bases (E.Ozarslan, H.E. Assemlal), and various regression and dictionary learning approaches. The challenge will provide a comparison both of different types of model, to see which predict the signal best, as well as which models within each category perform most strongly.

Time and location

Thursday, 16 April 2015 (8.30am-12.15pm)
Identifier: ThAT4
Location: Salon G & F

Contact Us

Daniel Alexander
Uran Ferizi
Torben Scheider
Benoit Scherrer
via e-mail:

The Complete Experiment Protocol

The full protocol consists of 48 HARDI shells, each with a different b-value arising from a different, unique combination of:

  • gradient strength |G| ~ {60, 100, 200, 300}mT/m;
  • pulse width δ = {3, 8}ms;
  • pulse duration Δ = {22, 40, 60, 80, 100, 120}ms;

The b-values thus range from 50 to 45,900 s/mm2, with effective diffusion time in the range 15 - 120ms. The protocol is designed to cover the measurement space as widely as possible. The raw images, consisting of 4mm sagittal slices, have voxels with in-plane dimensions of 2mm x 2mm. The set of "blip-up, blip-down" directions in each HARDI shell comes from a 45-direction evenly-distributed point set taken from Camino and acquired in both positive and negative direction to allow for eddy current distortion correction. The data was pre-processed with FSL using fnirt and eddy to correct for subject motion and eddy current distortions. The data for each voxel contains all processed "up/down" measurements and 10 interleaved b=0, giving 100 measurements in each shell.
(NOTE: You will actually find 12 more B0s in the files provided than stated. This is because every one of the TE-specific sets had 401 measurements originally: one extra b=0 measurement preceded every 400=4x(90DW+10B0) acquisitions. Because we keep out 100 measurements for assessment, there is an extra B0 preceding every 300 measurements instead.)

Training Data

The challenge data set, however, consists of two ROIs only, each containing 6 voxels. One comes from the middle of the genu in the corpus callosum, where the fibres are reasonably straight and coherent, while the other from the fornix, where the configuration of the fibres is more complex.
To obtain the training data, click
here for the genu and here for the fornix signal. Here is the corresponding protocol. All three are text files.

Each signal dataset contains 36 shells, or 3,612 diffusion and non-diffusion weighted measurements. The protocol's' columns specify the gradient direction (columns 1 to 3), the gradient strength in T/m (column 4), Δ in seconds (column 5), δ in seconds (column 6), and the echo time TE also in seconds (column 7). The repetition time is 1 second.

Here is a plot showing the b0-normalised signal in each of the genu's 48 shells against the cosine of the angle between the gradient direction and the normal to the sagittal slice (labelled n):

Test Data

To enter the challenge, you need to use your model trained on the data provided above to predict the missing 12 shells for each of the six voxels. Here is the specification of the imaging parameters for each measurement in the missing shells.
Click here to obtain the protocol for the unseen signal. The file format is the same as for the training data.

Note: the data has variable TE, so you may wish to estimate T2 beforehand from the b=0 measurements, or fit it together with your DW measurements.

Evaluating the Quality of Signal Prediction

The metric for determining how well the model predicts the unseen data is the sum of least squares of deviances of the predicted signal from the raw/hidden signal, but corrected for Rician noise, as used in the works quoted above (Panagiotaki et al.; Ferizi et al. (MICCAI'13)')):

The SNR of b=0 images is about 35 at TE=49ms and 6 at TE=152ms (with σ=8 for a TE-independent "thermal noise" "signal).

Submitting Your Work

To enter the contest, you need to submit two things:

  • A text file, similar to that of the training set, with your predictions for the missing signals of the six voxels. Please keep the order of the testing scheme file, as we will evaluate the prediction score automatically from this file.
  • A maximum one-page 12-pt-font pdf-format abstract describing your model. Please give it a recognisable name and provide enough information to understand the type of model it is, how many parameters it has, how you go about estimating the parameters from the data, and how you predict the unseen data. You do not need to describe the results in the abstract, as we get that from the data file above.

Email your entry to . For your abstract to be included in the workshop records/proceedings, and be considered for presentation at the workshop, we will need it by the closing date for the contest, March 15th.

We will however accept later entries and updates, up to March 30th, for inclusion in the final results we will post on the website and for contributions to the journal submission on the challenge (see below).

Journal Paper Collaboration

We plan to collate the results of the challenge into a journal paper detailing the findings. We will aim for NeuroImage for this paper and will add entrants whose work is included as co-authors. We may have to limit the number of co-authors from each team, but will certainly include at least the lead entrant. We will need to figure out the precise details once we have the full list of entries.

We look forward to seeing your entries and have fun with the competition!