Data

The dataset consists of 3 Tesla head MR images of 30 unrelated healthy subjects from the Human Connectome Project (HCP) dataset of healthy volunteers [2].

For each subject, the following data is available:

T1-weighted (T1w) MR image volume, not skull-stripped (but defaced for anonymisation [3]), with a bias field correction
T2-weighted (T2w) MR image volume, processed the same way as the T1w image
Both modalities in native T1w subject-space
The ground truth label map and brain mask in native subject-space
Affine transformation to align the images to the atlas (see below)

The ground truth labels are generated by FreeSurfer 5.3 (e.g., [4]) and are not manual expert annotations. As you will see when opening some example label maps, the automated labelling is imperfect. This is a common problem in the MIA domain; often, real expert annotations are sparse, and a "silver-standard" ground truth has to be used.

The data set used with MIA Lab is shared with students registered at Uni Bern. External users/students will have to access this data directly from the HCP webpage. Alternatively, most of this analysis works also on the Medical Segmentation Decathlon data sets - the hippocampus task is particularly easy and lightweight.

Atlas

The MR image and label files with the mni prefix are registered to the MNI152 atlas using nonlinear FNIRT.

T1-weighted atlas image: mni_icbm152_t1_tal_nlin_sym_09a.nii.gz
T2-weighted atlas image: mni_icbm152_t2_tal_nlin_sym_09a.nii.gz
Brain mask: mni_icbm152_t1_tal_nlin_sym_09a_mask.nii.gz

Add these files to the ./data/atlas/ directory.

Random forest toy example

To get a feeling of what a random forest - the type of machine learning classifier used to classify voxels in the brain tissues at interest - does, toy example data is provided. The toy example data files in the data directory (exp1_n2.txt, …) are taken from the Sherwood library [1].

References

[1] Microsoft Research, Sherwood C++ and C# code library for decision forests, 2012. [Online]. http://research.microsoft.com/en-us/downloads/52d5b9c3-a638-42a1-94a5-d549e2251728/

[2] Van Essen, D.C., Smith, S.M., Barch, D.M., Behrens, T.E., Yacoub, E., Ugurbil, K. and Wu-Minn HCP Consortium, 2013. The WU-Minn human connectome project: an overview. Neuroimage, 80, pp.62-79. Accessed from: http://www.sciencedirect.com/science/article/pii/S1053811913005351

[3] Milchenko, M. and Marcus, D., 2013. Obscuring surface anatomy in volumetric imaging data. Neuroinformatics, 11(1), pp.65-75. Accessed from: https://link.springer.com/article/10.1007/s12021-012-9160-3

[4] Fischl, B., Salat, D.H., Busa, E., Albert, M., Dieterich, M., Haselgrove, C., Van Der Kouwe, A., Killiany, R., Kennedy, D., Klaveness, S. and Montillo, A., 2002. Whole brain segmentation: automated labeling of neuroanatomical structures in the human brain. Neuron, 33(3), pp.341-355. Accessed from: http://www.sciencedirect.com/science/article/pii/S089662730200569X