Download

Thank you for your interest in the Calgary-Campinas-359 (CC-359) dataset. This dataset is constantly being updated. The dataset is now available to download at the Canada Open Neuroscience Platform (CONP) Portal! Updates to the dataset are initially be posted here, and subsequently added to the CONP Portal.

Download options:

Multi-vendor, multi-field-strength T1-weighted brain reconstructed data

This dataset is intended for developing data-driven automatic segmentation methods. At the moment, this dataset is composed of healthy brain images along with their brain masks "silver-standards" generated both using the STAPLE algorithm [1] and using supervised classification. We also provide white-matter and hippocampus "silver-standards" generated using STAPLE. We are prioritizing "silver-standards" generated with STAPLE as opposed to machine learning techniques to avoid having to (re-)train algorithms for every new structure that we include in the dataset, and also because results for STAPLE and machine learning seem to be within inter-rater variability.

The dataset is organized as follows:

  • Root folder:

    • Original: <id>_<vendor>_<field>_<age>_<gender>.nii.gz (359 volumes)


  • Skull-stripping-masks: ("silver-standards" generated from HWA, ROBEX, BET, OPTIBET, ANTs, BSE, MBWSS, BEaST)

    • "Silver standard" STAPLE probability mask: <id>_<vendor>_<field>_<age>_<gender>_staple.nii.gz (359 masks)

    • "Silver standard" binary mask: <id>_<vendor>_<field>_<age>_<gender>_ss.nii.gz (359 masks)

    • Manual segmentation: <id>_<vendor>_<field>_<age>_<gender>_manual.nii.gz (12 manual segmentations)

    • Manual segmentation by fixing BEaST mask: <id>_<vendor>_<field>_<age>_<gender>_man_julia.nii.gz (12 manual segmentations)

    • Manual segmentation by fixing 3D Watershed: (pending)


  • Hippocampus masks ("silver-standards" generated from FreeSurfer, FSL, Hipodeep)

    • "Silver standard" STAPLE probability mask: <id>_<vendor>_<field>_<age>_<gender>_hippocampus_staple.nii.gz (357 masks; CC0303 and CC0309 failed)

    • Hippocampus subfields computed with FreeSurfer: <id>_<vendor>_<field>_<age>_<gender>_*h_subfields.nii.gz (*: l for left hemisphere or r for right hemisphere)

Raw MR Data

Single-channel Coil Data (~8 GB uncompressed)

We are providing 35 fully-sampled (+10 for testing) T1-weighted MR datasets acquired on a clinical MR scanner (Discovery MR750; General Electric (GE) Healthcare, Waukesha, WI) . Data were acquired with a 12-channel imaging coil. The multi-coil k-space data was reconstructed using vendor supplied tools (Orchestra Toolbox; GE Healthcare). Coil sensitivity maps were normalized to produce a single complex-valued image set that could be back-transformed to regenerate complex k-space samples. You can see this data as a 3D acquisition in which the inverse Fourier Transform was applied on the readout direction, which essentially allows you to treat this problem as a 2D problem while at the same time undersampling on two directions (slice encoding and phase encoding). The matrix size is 256 x 256.

We are providing the train and validation sets. The data are split as follows:

  • Train: 25 subjects - 4,524 slices

  • Validation: 10 subjects - 1,700 slices

  • Test: 10 subjects - 1,700 slices (not provided, it will be used to test the model you submit)

This synthetic single-coil data is meant as proof of concept for assessing the ability of using Deep Learning for MR reconstruction. If you already have experience with that, we suggest that you go straight to the multi-channel coil reconstruction challenge, which is a more realistic scenario. We will not be adding more data to the single-channel coil track.

Multi-channel Coil Data (~219 GB uncompressed)

We are providing 167 three-dimensional (3D) , T1-weighted, gradient-recalled echo, 1 mm isotropic sagittal acquisitions collected on a clinical MR scanner (Discovery MR750; General Electric (GE) Healthcare, Waukesha, WI). The scans correspond to presumed healthy subjects (age: 44.5 years +/- 15.5 years [mean +/- standard deviation]; range: 20 years to 80 years). Datasets were acquired using either a 12-channel (117 scans) or a 32-channel coil (50 scans). Acquisition parameters were TR/TE/TI = 6.3 ms/ 2.6 ms/ 650 ms (93 scans) and TR/TE/TI = 7.4 ms/ 3.1 ms/ 400 ms (74 scans), with 170 to 180 contiguous 1.0-mm slices and a field of view of 256 mm x 218 mm. The acquisition matrix size for each channel was Nx x Ny x Nz = 256 x 218 x [170,180]. In the slice-encoded direction (kz), data were partially collected up to 85% of its matrix size and then zero filled to Nz= [170,180]. The scanner automatically applied the inverse Fourier Transform, using the fast Fourier transform (FFT) algorithms, to the kx-ky-kz k-space data in the frequency-encoded direction, so a hybrid x-ky-kz dataset was saved. This reduces the problem from 3D to two-dimensions, while still allowing to undersample k-space in the phase encoding and slice encoding directions. The partial Fourier reference data were reconstructed by taking the channel-wise iFFT of the collected k-spaces for each slice of the 3D volume and combining the outputs through the conventional sum of squares algorithm. The dataset train/validation/test split is summarized in the table below .Relevant information

      • Healthy subjects (age: 44.5 years ± 15.5 years; range: 20 years to 80 years).

      • Acquisition parameters are either: TR/TE/TI = 6.3 ms/2.6 ms/650 ms and TR/TE/TI = 7.4 ms/3.1 ms/400 ms

      • Average scan duration ~341 seconds

      • Only the undersampled k-spaces for R=5 and R=10 are provided for the test set


HYPOTHALAMUS SEGMENTATION BENCHMARK

Our benckmark is composed by a total of 1381 hypothalamus masks (manual and "silver standard") from four different datasets: IXI, CC359, OASIS, and MiLI, being the last one a new dataset, created for this benchmark.

We provide here 452 images from MiLI and 1343 masks from the mentioned datasets. We have 38 expert annotations hidden for benchmarking purposes.

MiLI, MICLab-LNI Initiative, comprises 452 T1-weighted MRI subjects, being 317 controls, and 135 patients with inherited ataxias and motor neuron diseases. The average age of the subjects is 36.14 years, with 212 declared male and 240 female. The images were acquired at the Hospital of the University of Campinas, and their usage is in agreement with the local ethics committee. All subjects underwent an MRI scan on a 3T Philips Achieva scanner (Philips, Best, The Netherlands) using standard 8-channel head coils. To segment the hypothalamus, we acquired 3D high-resolution T1 volumetric images of the brain with sagittal orientation, voxel matrix 240x240x180, voxel size 1x1x1mm3, TR/TE 7/3.201ms, and flip angle 8°.

THALAMUS BENCHMARKING


From the 1023 subjects selected from the HCP dataset to form this dataset, we provide the following split: the training set is composed of 963 subjects and contains only silver standard masks. The remaining 60 are manually segmented, with 20 having hidden labels for the benchmark hidden test.

Contact information

License

Our datasets are distributed under a Creative Commons Attribution-NoDerivatives 4.0 International Public License.


References:

[1] Warfield, S.K., Zou, K.H. and Wells, W.M., 2004. Simultaneous truth and performance level estimation (STAPLE): an algorithm for the validation of image segmentation. IEEE transactions on medical imaging, 23(7), pp.903-921.


Updated: 16 December 2021