Due to the large number of scans being collected in ICBM, it is important that vectorization be carried out automatically (algorithmically) as opposed to manually (which is still routine in neuronal morphology studies).
The basic algorithm we have designed and implemented to vectorize angiographic data is as follows. The volume is seeded in bright hotspots (typically along the Circle of Willis and positive voxels are searched through an iterative 26-neighbor strategy.
The Basic Vectorization Process: Seeding
The resulting images are then thresholded to eliminate spurious signals and artifacts (second to third panel). Finally, the angiographic skeleton (rightmost panel) is extracted by thinning (or erosion), which provides as well an estimate of branch diameter.
Several critical issues were examined. The first concerned the completeness of scanned sets. In most clinical neuroimaging applications, the acquisition of angiography data is typically limited to the most ventral regions. Several of the initial data sets we received were no exception to this protocol.
The Basic Vectorization Process: Flood Filling
More recent data sets received at the end of this second year were based on a modified scanning protocol, which includes the whole rostro-ventral extent.
A second critical aspect in the reconstruction quality is image resolution (voxel size). Data set #6, acquired at higher resolution than data set #2 yielded cleaner and more reliable digitizations.
The Basic Vectorization Process: Threshholding
We are now adapting the algorithms to the new protocol of phase contrast, which should allow the discrimination of veins and arteries. We also are focusing on fine-tuning the topology of the reconstructed trees (i.e., branch pruning and connectivity). The "gold standard" for this operation will be a limited set (N=1-2) of manually reconstructed volumes, to be carried out on forthcoming high-resolution phase contrast data sets.
The Basic Vectorization Process: Skeletonization