For every subject we provide additional information along with the images. These are:
- a text file with the 3D voxel coordinates of the centre of mass of the aneurysms and the maximum radius of the aneurysm
- a label image with labels:
- 0 = Background
- 1 = Untreated, unruptured aneurysm
- 2 = Treated aneurysms or artefacts resulting from treated aneurysms
The aim of the challenge is to automatically detect or segment untreated, unruptured aneurysms (label 0 vs label 1). Hence, we do not require methods to detect treated aneurysms and so provide a rough mask for them (label 2). Label 2 will be ignore during evaluation.
Detailed description of provided training data is provided here.
The protocol for aneurysm annotation was developed by an interventional neuro-radiologist, with more than 10 years of experience in the field . On each axial slice of the TOF-MRA a contour around the outline of the aneurysm was drawn. The annotations were always drawn to be from the level of the neck to the dome of the aneurysm. The neck corresponds to the opening of the aneurysm from the parent vessel. The dome is the furthest part of the aneurysm from the parent vessel. None of the parent vessel was included in the annotation. During annotation, the raters had access to the structural image and a radiologist report made at the time of the scan. The radiologists’ reports indicated the rough location and size of the aneurysm.
Aneurysms were diagnosed and located in all images as part of clinical routine. The experienced interventional neuro-radiologist trained a second rater with extensive experience in medical image analysis and the annotation software.
Once this second rater was on par with the first rater , the rater annotated all images in the dataset. Finally, the first and second rater assessed the full dataset together and made any required modifications in consensus. This consensus annotation was used to produce the binary masks as the official ground truth data set.
The contours were converted to binary masks, including all voxels whose volume >50% was included within the manually drawn contour. Background received label 0 and untreated, unruptured aneurysms label 1.
Treated aneurysms (or artifacts resulting from treated aneurysms) were converted to binary masks as well, receiving label 2. These masks were dilated by 1 pixel in-plane (with a 3x3x1 kernel). In case of overlap between labels 1 and 2, label 1 was assigned.
Inter observer agreement
A subset of the training cases were also manually segmented by the experienced observer. The final inter observer agreement between the consensus annotations and the experienced observer will be announced as soon as possible.