Figure 1.
Workflow of the radial symmetry transform.
An example slice of a 3.0 T T2*-weighted image with a microbleed is shown in A. B shows a zoom-in of the microbleed shown in A. An intermediate step of the radial symmetry transform is shown in C, where each voxel contributes to some neighboring voxel. If multiple arrows point towards the same target voxel, this target voxel receives a high radial symmetry value. The output of the radial symmetry transform is shown in D, displaying a large value at the center voxel of the microbleed.
Figure 2.
Results of the radial symmetry transform for detecting microbleeds.
Appropriate values for and
were determined on five randomly selected participants with microbleeds. The figures show the results of these values on all 72 participants. A) Number of detected potential microbleed locations per participant (n = 72), visualized as a function of
and
. Potential microbleeds need to be censored by a human rater to identify true microbleeds (on average 2 per participant) and reject false positives that remained after the 3D and 2D radial symmetry transform. B) Sensitivity of the radial symmetry transform on the visual ground truth, visualized as a function of
and
. C) This figure shows the relationship between sensitivity and the number of potential microbleed locations per participant, where each dot is a combinations of
and
. Depending on the preferred sensitivity, there is an optimal combination of
and
with the lowest number of locations. These optimal combinations are annotated by the solid line. Three combinations (A, B, and C, annotated from left to right with squares), with moderate, good, and high sensitivity, were selected for inspection by a human rater.
Table 1.
Results of the radial symmetry transform after censoring potential microbleed locations per participant.
Table 2.
Sensitivity of manual and semi-automated rating.