Task 1: Segmentation of Enlarged Perivascular Spaces¶

On this page we introduce and explain task 1 in more detail. Please see the description page, data page and the evaluation page for more information about this task.

Goal¶

The goal of this task is to develop an automated method to segment enlarged perivascular spaces (PVS) in MRI scans. Annotating enlarged PVS is extremely time-consuming and cumbersome, so there is an important trade-off between annotation "strength" (counts vs segmentations) and how many scans can be annotated. The training set consists of cases with different annotation types in various portions of the brain. The submitted automated methods will be applied on the hidden test set and should output segmentations of enlarged PVS for the full brain. The predicted segmentation masks should be continuous and will be thresholded at 0.5 during evaluation to obtain binary segmentation masks (0: background, 1: enlarged PVS). 

Data¶

See the  data page.

Annotation type¶

This table indicates the number of cases per annotation type in the training set:

This table indicates the number of cases in the test set:

Counts¶

All cases contain a csv file indicating the number of enlarged PVS per label in the region masks. The file has the following columns: index, value (=label in region mask), count (=number of enlarged PVS), min_slice (=lower slice in region mask), max_slice (=upper slice in region mask). Note that slice numbers are 0-indexed. For 28 of the 40 cases this is the only annotation type that is provided.

Segmentations¶

For 12 of the 40 cases segmentation masks have been provided. For the RSS cases only 1 segmentation mask is available from 1 rater, for SABRE 2 segmentation masks are provided as the cases have been annotated by 2 raters. An example of segmentations in the centrum semiovale:

Region Masks¶

The region masks indicate which parts of the scans have been annotated. The values in the region mask indicate the label of this portion of the image (slab / slice / region) and correspond to the labels in the count csv files. 

Slabs (SABRE)¶

For all SABRE cases segmentations have been provided in slabs (=stacks of slices). The region masks indicates which slices have been annotated:

Regions and Slices (RSS)¶

For all RSS cases annotations have been done in the most clinically relevant regions, namely in the centrum semiovale (white matter), basal ganglia, hippocampi, midbrain. The midbrain and hippocampi were fully annotated for all cases. As the centrum semiovale and the basal ganglia are larger regions, only 1 axial slice per region was annotated. For the basal ganglia the slice showing the anterior commissure was annotated and for the centrum semiovale the slice 10 cm above the lateral ventricle (for all cases). For the cases with segmentations 1 additional axial slice of the basal ganglia was annotated and 1 axial slice containing white matter, these slices were chosen randomly.

The region masks indicates which part of the scans have been annotated (see example below for a case with segmentations). In the count csv file the number of enlarged PVS is indicated per label of the region mask as well as the slice numbers. 

Since the update of the training data (see announcements April 2021) the region labels across all RSS cases are consistent. The following labels are used:

* For the segmentation cases both region labels are used as 2 slices are annotated, for the count cases only 1 of the labels is used as only 1 slice is annotated

Evaluation¶

Methods will be evaluated on segmentation, volume estimation, detection and count estimation. See the  evaluation page.