The 3D cardiovascular magnetic resonance (CMR) images were acquired during clinical practice at Boston Children’s Hospital, Boston, MA, USA. Cases include a variety of congenital heart defects. Some subjects have undergone interventions.
Imaging was done in an axial view on a 1.5T scanner (Phillips Achieva) without contrast agent using a steady-state free precession (SSFP) pulse sequence. Subjects breathed freely during the scan, and ECG and respiratory-navigator gating were used to remove cardiac and respiratory motion (TR = 3.4 ms, TE = 1.7 ms, α = 60˚). Image dimension and image spacing varied across subjects, and average 390 x 390 x 165 and 0.9 x 0.9 x 0.85 mm, respectively, in the full-volume training dataset.
Manual segmentation of the blood pool and ventricular myocardium was performed by a trained rater, and validated by two clinical experts. Segmentations were done in an approximate short-axis view and then transformed back to the original image space (axial view). Manual segmentation was done considering all three planes, but the quality of the segmentation in the short-axis view was the deciding factor.
The blood pool class includes the left and right atria, left and right ventricles, aorta, pulmonary veins, pulmonary arteries, and the superior and inferior vena cava. Vessels (except the aorta) are extended only a few centimeters past their origin: this is because vessels that are too long are disruptive when the 3D heart surface models are used for surgical planning. The myocardium class includes the thick muscle surrounding the two ventricles and the septum between them. Coronaries are not included in the blood pool class, and are labeled as myocardium if they travel within the ventricular myocardium.
Update: As requested, we have released cropped and short-axis images for each subject. The approximate short-axis view was manually defined using the mitral valve and apex only, and so there is residual rotation around this axis that is not consistent across subjects. Challenge participants therefore have three options for which data to use for training/testing: (1) The axial, full-volume images, i.e. the original acquisition (2) Cropped axial images (3) Cropped images that have been transformed into an approximate short-axis view
If you use this data as part of a research paper, please cite the following paper:
- D.F. Pace, A.V. Dalca, T. Geva, A.J. Powell, M.H. Moghari, P. Golland, “Interactive whole-heart segmentation in congenital heart disease”, Medical Image Computing and Computer Assisted Interventions (MICCAI 2015), Lecture Notes in Computer Science; 9351:80-88, 2015.
All challenge data and the results of the participating groups will be released openly. All data will remain publically available after the challenge to facilitate meta-analyses.
The challenge system and data are unfortunately no longer available.
Please watch this space, we are working on releasing additional datasets with more extensive labeling!
The training and testing data is hosted on the HVSMR 2016 Challenge system.
To download the training data without registering online, you can also use the following links:
- Axial, full-volume training data: images, labels
- Axial, cropped training data: images, labels
- Short-axis, cropped training data: images, labels
To download the testing data without registering online, you can also use the following links (in the testing phase, ground truth labels are not made public):
- Axial, full-volume training data: images
- Axial, cropped training data: images
- Short-axis, cropped training data: images
Awards for the best segmentation algorithms will be provided by: