Parallelized Segmentation of CT-Angiography datasets using CUDA

Segmentation of CT-Angiography datasets is an important and difficult task. Several algorithms and approaches have already been invented and implemented to solve this problem. In this work, we present automatic algorithms for the segmentation of these CTA datasets, implemented in CUDA, and evaluate our results regarding speed and error rates. Starting with local approaches like thresholding we proceed to global, object-based algorithms, like region growing and a newly developed algorithm based on dual energy CT scans (DECT), the XOR-Algorithm, presented by Karimov et al.[6] A limitation of using graphics hardware is the restricted amount of memory, which led us to use a slab-based processing approach (see section 5.3). The requirement of this work was a complete GPU implementation. But since not every task is appropriate for parallelizing, it was necessary to use iteratively parallel algorithms. This strategy though introduced speed problems that had to be analysed and were partly solved. This work presents the principle of these GPU methods and compares them to their CPU counterparts. In the end, the quality of each algorithm is analysed and they are compared against each other, in order to find an acceptable completely automatic segmentation algorithm for distinguishing between different types of tissues (e.g. vessels, bones, soft tissue, ...).