About
Introduction
The goal of this challenge is to compare 3D MS lesion segmentation techniques. General lesion segmentation methods tailored for multiple sclerosis segmentation will be accepted for the competition as well. Algorithms for the competetion should be fully automatic methods. The performance of the MS lesion segmentation methods will be evaluated using a set of comprehensive measures.
Developers of MS lesion segmentation techniques from both academia and industry are welcomed to join the contest.
MS lesion data
MS lesion MRI image data for this competition was acquired seperately by Children's Hospital Boston and University of North Carolina. UNC cases were acquired on Siemens 3T Allegra MRI scanner with slice thickness of 1mm and in-plane resolution of 0.5mm. To ease the segmentation process all data has been rigidly registered to a common reference frame and resliced to isotrophic voxel spacing using b-spline based interpolation. Pre-processed data is stored in NRRD format containing an ASCII readable header and a separate uncompressed raw image data file. This format is ITK compatible. Full documentation is available
here.
MS lesion training and testing data
Data used for this workshop is composed of 54 brain MRI images and represents a range of patients and pathology which was acquired from Children's Hospital Boston and University of North Carolian. Data has initially been randomized into three groups: 20 training MRI images, 24 testing images for the qualifying and 8 for the onsite contest at the 2008 workshop. The downloadable online database consists now of the training images (including reference segmentations) and all the 32 combined testing images (without segmentations). The naming has not been changed in comparison to the workshop compeition in order to allow easy comparison between the workshop papers and the online database papers. One dataset has been removed (UNC_test1_Case02) due to considerable motion present only in its T2 image (without motion artifacts in T1 and FLAIR). Such a dataset unfairly penalizes methods that use T2 images versus methods that don't use the T2 image.
Currently all cases have been segmented by expert raters at each institution. They have significant intersite variablility in segmentation.
The content of this website is copyrighted © Martin Styner, Simon Warfield,Wiro Niessen, Theo van Walsum, Coert Metz, Michiel Schaap, Xiang Deng, Tobias Heimann, and Bram van Ginneken. This work is supported by the UNC Neurodevelopmental Disorders Research Center HD 03110 and NIH Roadmap for Medical Research, National Alliance for Medical Image Computing, Grant U54 EB005149-01.