The Challenge

Many algorithms for segmenting brain structures in MRI scans have been proposed over the years. Especially in such a well-established research area, there is a tremendous need for fair comparison of these methods with respect to accuracy and robustness. Publicly available databases like the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and the Internet Brain Segmentation Repository (IBSR) are great initiatives and promote the use of publicly available data for automatic segmentation algorithm comparison, but either lack full manual segmentations (ADNI) or provide low-field (1.5T) single sequence (T1-weighted) MRI data (IBSR). Although there is an increasing awareness of the importance of comparing different algorithms on the same data, many methods are still compared to previous versions of the same type of algorithm on privately held data. This complicates the choice for a certain brain segmentation method among a wide variety of available methods. Furthermore, high-field (3T) multi-sequence MRI data is increasingly available in the clinic, while some available methods were developed for 1.5T single sequence T1-weighted MRI data. Comparing these methods to multi-sequence approaches on high-field (3T) data is of great interest to the medical image processing community.

The aim of this challenge is to compare (semi-)automatic algorithms for segmentation of gray matter, white matter and cerebrospinal fluid on 3T MRI scans of the brain. Since participants to the challenge apply their own methods to the data, there is no bias toward one particular method, and parameters are optimally tuned to achieve the best possible performance. For the challenge, twenty fully annotated multi-sequence (T1-weighted, T1-weighted inversion recovery and FLAIR) 3T MRI brain scans are available. More information on the data can be found here. We encourage both authors of new methods as well as authors of already published brain segmentation methods to participate in the challenge and tune their algorithms in the best possible way on the training data, to enable a fair comparison of the methods. We welcome both multi- and single-sequence (only T1-weighted scan) approaches.

If you wish to participate in the MRBrainS challenge, read the details on this page and follow the steps below:

  1. Register on this website, sign the MRBrainS13 confidentiality agreement and e-mail it to mrbrains13@isi.uu.nl.
  2. After you have received a confirmation e-mail, you can download the 5 training datasets and 15 test datasets on the download page.
  3. Use the training data to fine-tune your segmentation algorithm.
  4. Apply your segmentation algorithm to the test datasets.
  5. Submit your results in the format described here together with a description of your method, your evaluation results will be sent to you by e-mail, and will be published on this website in the results section.

We kept the segmentation task as simple as possible, to allow more algorithms to participate. Participants should segment gray matter, white matter and cerebrospinal fluid. Algorithms that are able to segment substructures, such as cortical gray matter and basal ganglia, should merge these results and label them as gray matter. This is described in further detail in the Test Data section. In the manual segmentations provided for training, the substructures are labeled separately to allow algorithms that segment substructures to use these labels. Authors with algorithms that only use labels for gray matter, white matter and cerebrospinal fluid, should merge these labels themselves. This is described in further detail in the Training Data section. Both multi-sequence as well as single-sequence (using only the T1-weighted scan) approaches are welcome to participate. We will distinguish between these approaches in the results section.

Twenty fully annotated multi-sequence (T1-weighted, T1-weighted inversion recovery and FLAIR) 3T MRI brain scans are available. These scans have been acquired at the UMC Utrecht (the Netherlands) of patients with diabetes and matched controls (with increased cardiovascular risk) with varying degrees of atrophy and white matter lesions (age > 50).

For each patient, the following sequences are provided:

  1. T1_1mm: 3D T1-weighted scan (voxel size: 1.0mm x 1.0mm x 1.0mm)
  2. T1: 3D T1-weighted scan registered to the T2 FLAIR (voxel size: 0.958mm x 0.958mm x 3.0mm)
  3. T1_IR: Multi-slice T1-weighted inversion recovery scan registered to the T2 FLAIR (voxel size: 0.958mm x 0.958mm x 3.0mm)
  4. T2_FLAIR: Multi-slice FLAIR scan (voxel size: 0.958mm x 0.958mm x 3.0mm)

The Segmentation Task

We kept the segmentation task as simple as possible, to allow more algorithms to participate. Participants should segment gray matter, white matter and cerebrospinal fluid. Algorithms that are able to segment substructures, such as cortical gray matter and basal ganglia, should merge these results and label them as gray matter. This is described in further detail in the Test Data section. In the manual segmentations provided for training, the substructures are labeled separately to allow algorithms that segment substructures to use these labels. Authors with algorithms that only use labels for gray matter, white matter and cerebrospinal fluid, should merge these labels themselves. This is described in further detail in the Training Data section. Both multi-sequence as well as single-sequence (using only the T1-weighted scan) approaches are welcome to participate. We will distinguish between these approaches in the results section.

Twenty fully annotated multi-sequence (T1-weighted, T1-weighted inversion recovery and FLAIR) 3T MRI brain scans are available. These scans have been acquired at the UMC Utrecht (the Netherlands) of patients with diabetes and matched controls (with increased cardiovascular risk) with varying degrees of atrophy and white matter lesions (age > 50).

For each patient, the following sequences are provided:

FileDetails
T1_1mm3D T1-weighted scan (voxel size: 1.0mm×1.0mm×1.0mm)
T13D T1-weighted scan registered to the T2 FLAIR (voxel size: 0.958mm×0.958mm×3.0mm)
T1_IRMulti-slice T1-weighted inversion recovery scan registered to the T2 FLAIR (voxel size: 0.958mm×0.958mm×3.0mm)
T2_FLAIRMulti-slice FLAIR scan (voxel size: 0.958mm×0.958mm×3.0mm)

All scans are bias corrected, and scans 2-4 are aligned. This is done to eliminate the influence of different registration and bias correction algorithms on the segmentation results. The manual labeling is done on scans 2-4. The 3D T1-weighted scan is provided for algorithms that prefer to use 3D high-resolution data. However, the final results of your segmentation method should be aligned with scan 2-4, since the ground truth is only available for these scans! The face of the patient is cut out of the 3D T1-weighted scan for the purpose of anonymization. Some of the T1-weighted IR images contain artifacts at the bottom of the scan. These artifacts occur often in clinical scans, and it is therefore interesting to know how automatic segmentation methods that use these scans perform under these circumstances.