high performance computing on graphics processing units: hgpu.org

Image segmentation is the task of splitting a digital image into one or more regions of interest. It is a fundamental problem in computer vision and many different methods, each with their own advantages and disadvantages, exist for the task. Image segmentation is a particularly difficult task for several reasons. Firstly, the ambiguous nature of splitting up images into objects of interest requires a trade off between making algorithms more generalized or having many user specified parameters. Secondly, image artifacts such as noise, inhomogeneity, acquisition artifacts and poor contrast, are very difficult to account for in segmentation algorithms without a high level of interactivity from the user. In this report, segmentation is discussed in a medical imaging context however the proposed algorithm could equally be used in general purpose segmentations. Segmented images are typically used as the input for applications such as classification, shape analysis and measurement. In medical image processing, segmented images are used for studying anatomical structures, diagnosis and assisting in surgical planning. Before computational algorithms existed, segmentation of medical images was a tedious process, performed by hand by clinical experts. This was a fairly accurate, yet slow process. These expert segmentations form the gold standard with which to validate algorithmic segmentations. As there is no single general solution to image segmentation problems, several techniques exist each of which has its own strengths and limitations. Some common techniques are thresholding based segmentation, region based segmentation (i.e. region growing), edge based segmentation and deformable active contour models (such the snakes and geodesic active contour models).