This title appears in the Scientific Report :
2013
GPU-accelerated Segmentation of High-resolution Human Brain Images Acquired with Polarized Light Imaging
GPU-accelerated Segmentation of High-resolution Human Brain Images Acquired with Polarized Light Imaging
High-resolution three-dimensional polarized light imaging (PLI) is an approach pursued by INM-1 (Institute of Neuroscience and Medicine, Forschungszentrum Jülich) to map nerve fibers and their pathways in human brains. The 100 micron thick sections of the cut post-mortem brain are imaged with a micr...
Saved in:
Personal Name(s): | Westhoff, Anna (Corresponding author) |
---|---|
Bücker, Oliver / Axer, Markus / Grotendorst, Johannes | |
Contributing Institute: |
JARA - HPC; JARA-HPC Strukturelle und funktionelle Organisation des Gehirns; INM-1 Jülich Supercomputing Center; JSC |
Imprint: |
2013
|
Conference: | Bernstein Conference on Computational Neuroscience 2013, Tübingen (Germany), 2013-09-24 - 2013-09-27 |
Document Type: |
Poster |
Research Program: |
Computational Science and Mathematical Methods |
Publikationsportal JuSER |
High-resolution three-dimensional polarized light imaging (PLI) is an approach pursued by INM-1 (Institute of Neuroscience and Medicine, Forschungszentrum Jülich) to map nerve fibers and their pathways in human brains. The 100 micron thick sections of the cut post-mortem brain are imaged with a microscopic device using polarized light. This way the birefringence of the myelin sheaths surrounding nerve fiber axons allows to extract a vector field of fiber tract orientations that form the basis for tractography.
The section is moved during the imaging within the microscope so that a mosaic of about 30x30 image tiles is created for a gross-histological human brain section. These up to 900 tiles per section have to be handled in the 3D-reconstruction of the brain. One of the most challenging and expensive tasks of the reconstruction pipeline is the registration of the sections among each other which is needed due to non-linear deformations occurring during the brain preparation. A way to accelerate this process is a segmentation of the image tiles which leads to black-and-white masks marking the brain and background pixels of the original tiles. Hence, all non-brain parts of the tiles can be ignored during the registration.
A region growing segmentation has been enhanced and specialised for the PLI data. The challenge to adapt this algorithm to the given dataset is to automize the choice of seeds needed as starting points for the growing process. Therefore, an automated method of seed determination has been developed. It uses statistics of the whole brain based on the image histogram. This approach leads to a minimal fixed amount of required manual input which is independent of the number of images to be segmented. The parallel software tool is developed to be executed on JUDGE, a supercomputer hosted by JSC (Jülich Supercomputing Centre, Forschungszentrum Jülich). Calculating parts of the algorithm on GPUs reduces the execution time by a factor of 10. |