- Make a prediction for each test data (54 data with 4752 slices in total).
- Encode your predictions in RLE for each slice of each test data in the format of the template file (provided with the test data).
- Upload your prediction (single .csv file) HERE for evaluation:
- All participants (or research teams) are allowed a maximum of valid 5 entries in the challenge.
The participants are required to produce segmentations of the left atrial (LA) cavity. For the .csv submission file, run-length encoding (RLE) on the pixels will be used to reduce the size of the files, and speed up the evaluation process. The RLE format is such that, instead of an exhaustive list of indices, pairs of values are uploaded containing the start position and the run length for each positive pixel region. An example would be ‘2 5’, where the pixels at positions 2,3,4,5 and 6 (5 in total starting from 2) are deemed positives by the algorithm. The encodings shall be in a space delimited format such that ‘2 5 9 3’ implies pixels at positions 2,3,4,5,6,9,10 and 11 are positives.
Since the challenge data comprises of 3D image data, participants are required to submit the RLE for each 2D slice of every GE-MRI along the z-axis (shortest-axis) individually. The format of the .csv submission file should contain only 2 columns in a comma delimited format. The first is the MRI ID and slice number for that specific MRI, the second is the RLE for each 2D slice for the LA cavity. An example is shown below. The slices for each MRI should be sorted in ascending order from the bottom slice to the top slice of the MRI. Since each individual slice of an MRI is 2-dimensional, the pixels are one-indexed, first from top to bottom then left to right, when encoded into RLE. An example is shown below.
*It is highly recommended participants use the template submission file (provided with the test data) as it includes all the rows necessary for prediction.
The organizers recognize 3D image segmentation tasks are not common and participants may come across various issues throughout the challenge. The template .csv files are available as a guide for submission. Sample python code is provided here to demonstrate how to load the data, how to encode the data using RLE and how to write to the submission file. We have also provided the RLE for all the training data labels in a separate .csv file so the participants can check their RLE algorithm against those of the training replica set. The organizers will remain active throughout the competition to answer questions and provide help in case of any issues met by the participants. Please post any questions in the Discussion Forum section for this challenge to benefit all participants.