Accelerated implementation of level set based segmentation

Abstract

An Open Computing Language implementation of a level set solver for 2D and 3D image segmentation tasks is presented. An adaptive time stepping algorithm is implemented using an optimised parallel reduction kernel to compensate for a loss of algorithmic parallelisation. For a 2D data set (256×256) the execution is accelerated by a factor of 20 in the adaptive case and 100 in the non-adaptive case compared to a cpu implementation, facilitating real time interactive parameter tuning. For a 3D data set (384×397×41) the acceleration factors are 200 and 270 for the adaptive and non-adaptive cases, respectively. Although a single iteration of the adaptive method is slower compared to the non-adaptive scheme, it automatically enforces the Courant,Friedrichs, Lewy condition and reduces the number of user-tuned parameters while safely allowing larger time steps. Open Computing Language optimisations and techniques are discussed