A work-efficient GPU algorithm for level set segmentation

Abstract

Figure 1: The progression of our algorithm while segmenting the brain matter in a 256 3 head MRI with a signal-to-noise ratio of 11. Our algorithm interactively computes this segmentation in 7 seconds -14× faster than previous GPU algorithms with no reduction in accuracy.Abstract We present a novel GPU level set segmentation algorithm that is both work-efficient and step-efficient. Our algorithm has O(log n) step-complexity, in contrast to previous GPU algorithms [Lefohn et al. 2004;Jeong et al. 2009] which have O(n) step-complexity. Moreover our algorithm limits the active computational domain to the minimal set of changing elements by examining both the temporal and spatial derivatives of the level set field. We apply our algorithm to 3D medical images ( Figure 1) and demonstrate that our algorithm reduces the total number of processed level set field elements by 16× and is 14× faster than previous GPU algorithms with no reduction in segmentation accuracy.Introduction Identifying distinct regions in images -a task known as segmentation -is an important task in computer vision and medical imaging. The GPU narrow band algorithm for level set segmentation can compute highly accurate segmentation results for noisy medical images and dramatically reduces computation times compared to optimized CPU implementations. However the GPU narrow band solver we tested took over 100 seconds converge on the brain matter in a 256 3 head MRI on an Nvidia GTX 280 ( Figure 2). This limitation constrains clinical applications and motivates our work-efficient algorithm.

Fast Sparse Level Sets on Graphics Hardware

“…Unfortunately, both methods have rather high memory requirements. However, Roberts et al [31] show that their method is both work and step efficient.…”

“…Both methods are more computationally involved than the pure level-set method, and thus a direct comparison with regard to efficiency is unfair. While not sparse, the recent methods of Roberts et al [31] and Klar [15] represent approaches to implement narrow-band approaches on the GPU. Unfortunately, both methods have rather high memory requirements.…”

“…Unlike our method, the GPU method of Roberts et al [31] is not sparse. Actually, the latter method has rather high memory requirements, which drastically limits the maximum resolution that can be achieved.…”

“…Note that the method is not based on tiles, but instead it works directly with sets of 3D locations. In [31], it is shown that the method is both work-efficient and step-efficient, and it is faster than Lefohn's method.…”

Medical image processing on the GPU – Past, present and future

“…Hagan and Zhao (2009) used a Lattice Boltzmann model (LBM) approach to solve level set-based segmentation, a GPU cluster was used to handle large medical volumes. More recent examples include combining level sets and CUDA for segmentation of CT and MRI data (Roberts et al, 2010;Sharma et al, 2010).…”