Thanks for your message!
The aim of Task 1 is to develop an automated method that segments enlarged perivascular spaces in the full brain based on MRI scans (the available sequences are T1w, T2w and FLAIR). So based on the MRI scans the method should output a semantic segmentation of the same size as 1 input scan. When thresholded at 0.5 this semantic segmentation should be a binary segmentation mask with voxels that are 0 if they belong to the background and 1 for enlarged perivascular spaces.
The training set consists of cases with different annotation types that can be used to develop/train this method. Participants can choose which annotation types they use and which cases. E.g. only cases with segmentations can be used, which means less cases can be used for training; or all cases can be used and only the counts (weakly supervised methods); or a combination of all annotation types can be used; etc.
The ground truth is the PVSSeg nifti, but only for the parts of the scan indicated in the region masks annotations are provided. There is no full ground truth given in the training set.
Background voxels are voxels that are not enlarged perivascular spaces, so anything that is not an enlarged perivascular space.
We recently added some extra information on this task and the available annotations on this page: https://valdo.grand-challenge.org/Task1/
Hope it is more clear now! Otherwise please let us know!