Judging

From now to February 2013
Each submission contains a set of water/fat-separated images, one for each case. For each available case, a fat-fraction (FF) map is calculated from the fat and water images. A reference FF map is pre-calculated for each case using an expanded data set with additional echo times compared to the download data and the following signal model.
Reference signal model

The signal model, neglecting B0 inhomogeneity and R2* decay, is given by
   
where water is assumed to be at 0 Hz, and fat is assumed to contain 6 spectral peaks at Δfp = -242.7060, -217.1580, -166.0620, -123.9078, -24.9093, 38.3220 Hz at 1.5T, with relative amplitudes ap = 0.0870, 0.6930, 0.1280, 0.0040, 0.0390, 0.0480.

Assumption on R2*: A single R2* constant is assumed per pixel.

The FF map is compared with the reference FF map within a predefined support mask. Fat fraction difference within +/- 0.1 is neglected. The number of matching voxels (i.e. absolute difference less than 0.1) is calculated with a score from 0 to 10000, with 10000 indicating complete agreement. A score of 0 is assigned to each missing case. (Note: it is better to submit an imperfect reconstruction than to leave it out.) The overall score is determined as the average across all cases.

From February 2013 onwards
The competition will be closed to the public. The top teams are invited to submit their algorithms (can be source code or executable) within two (2) weeks. The submitted algorithms will be applied by the Challenge organizers to an expanded set of data without case-by-case adjustments. The score is calculated as before; the processing time is verified to be under <30min per case; and a physician panel verifies the separation results. The selected teams may re-submit their algorithms as often as once per week to improve their score.