A total of 154 3D MRIs from patients with AF are used in for the purpose of this challenge. A large proportion of data were kindly provided by The University of Utah (NIH/NIGMS Center for Integrative Biomedical Computing (CIBC)), while the rest were from multiple other institutes. All clinical data have obtained institutional ethics approval. Each 3D MRI patient data was acquired using a clinical whole-body MRI scanner and contained raw the MRI scan and the corresponding ground truth labels for the left atrial (LA) cavity. The ground truths were manually segmented by experts in the field. The raw MRIs are in grayscale and the segmentation labels are in binary (255 = positive, 0 = negative). The dimensions of the MRIs may vary depending on each patient, however, all MRIs contain exactly 88 slices in the Z axis.
The dataset is split such that 100 patient data are used for training, and 25 patient data will be used for testing and evaluation. The participants will have access to all the MRIs and their respective labels (LA cavity mask) in the training set, and only the MRIs in the testing set.
To deter manual segmentation, the test data will be released 2 weeks prior to the end of the challenge such that participants should first develop their model on the training set, and then submit predictions for the test set within the 2-week window.
- All 100 training data for LA cavity segmentation released (data + mask)
2 weeks before competition deadline:
- 25 test data for LA cavity segmentation released (data only)
The files are arranged such that each individual file contains one patient data. For each data, the raw MRI “lgemri.nrrd”, the LA cavity segmentation “laendo.nrrd” are provided. “.nrrd” is a medical imaging file format, and can be read using various programming languages.
Once you have signed the Data Access Agreement Form, you will have access to the following:
Training Set (folder) – the entire training set (100 MRIs and LA cavity labels)
submission.csv – the submission file containing the MRI_ID_SliceNumber to predict for
train_labels.csv – the run-length encodings of the training images (for convenience and checking purposes)
sample_code.py – sample python code for the challenge to get participants started