high performance computing on graphics processing units: hgpu.org

In this work, a speed comparison between GPU-based CUDA k-NN implementation and the ANN implementation has been tested on three sets of medical imaging data. The results show that with higher dimensional data, CUDA-based k-NN approach could have up to two orders of magnitude of speed up. Otherwise, ANN would be a better implementation to use. Also, based on the work of CUDA k-NN, we present two new approaches to segment vasculature in mouse retina using large set of features. They performs better than directly implementing closest previous OCT-based approach. The method using all layer projections would show overall best result with more visible vessels and higher contrast.