CHALLENGE DATASETS
Data of patients with cerebral aneurysms without vasospasm were collected for diagnostic and treatment decision purposes.
The image data were acquired utilizing the digital subtraction AXIOM Artis C-arm system using a rotational acquisition time of 5 s with 126 frames (190° or 1.5° per frame, 1024 x 1024-pixel matrix, 126 frames). Post-processing was performed using LEONARDO InSpace 3D (Siemens, Forchheim, Germany). A contrast agent (Imeron 300, Bracco Imaging Deutschland GmbH, Germany) was manually injected into the internal carotid (anterior aneurysms) or vertebral (posterior aneurysms) artery. Reconstruction of a volume of interest selected by a neurosurgeon generated a stack of ~220 image slices with matrices of 256x256 voxels in-plane, resulting in an iso-voxel size of ~0.5 mm.
An experienced annotator has provided the segmentations. The windowing used for image annotation was not fixed. The users might have worked with different settings in the consecutive annotation sessions. The different settings might affect the assumed extensions of the aneurysms. An experienced neurosurgeon checked all segmentations. Because the original goal in segmentation aimed at using the surface meshes for CFD, the segmentation masks might be smoother than the actual vessel surface. Labels are available for all datasets.
The criterium for ruptured aneurysms is subarachnoid hemorrhage (SAH) seen on CT scans (most often) or MRI scans. In cases of normal CT or MRI scans but with symptoms suspicious of a subarachnoid hemorrhage, typical findings in cerebrospinal fluid (lumbar puncture) prove a subarachnoid hemorrhage. In cases of multiple aneurysms, the localization (asymmetry) of the SAH points to the ruptured aneurysm. If there is no asymmetry of the SAH, the more proximal or greater or, the more irregularly shaped aneurysm is considered to be the ruptured one. In rare cases, a xanthochrom parenchymal „halo“ around the aneurysm seen in surgery proves the history of a (small and past) hemorrhage originating from this aneurysm.
In addition to the image data, mask and stl-files are provided for the segmented aneurysms. The rupture information is stored in a table listing the aneurysms together with their rupture state (58% not ruptured vs 42% ruptured).
The average volume per aneurysm is 0.391 ml, meaning that for image volumes between 280 and 2350 ml, there is a strong imbalance between foreground and background, which the participants have to consider when designing the preprocessing and training setup.
Data for Task 1
A training case consists of an image dataset showing a contrast-enhanced cerebrovascular vessel tree. Furthermore, segmentation masks (stl- and image files) are provided, so that the bounding boxes for the aneurysms can be calculated.
The test cases for generating results to be uploaded contain only image data.
The subset of the data to be used as training data comprises 109 datasets with 127 annotated aneurysms. It is released to the participants and can be freely partitioned e.g. for cross validation.
