Denoising CT Enterography on GPU

CT enterography is a method of contrast enhanced diagnosis of small intestine from computed tomography examination. However, unlike the bones or other significant structures, automatic or semiautomatic examination of small intestine is almost impossible without the proper use of a denoising algorithm. We cannot wait for hours for a simple denoising algorithm in medical praxis, but current GPUs enable the use of such computationally intensive algorithms like nonlocal means even on volumetric data. We discuss the implementation of such algorithms with different modifications in respect to the performance and suitability to run on a GPU. Introduction For diagnosis and treatment preparation of small intestine problems, such as Crohn’s disease, exists a method called CT enterography (Federle, 2007) (Paulsen et al., 2006). It is a noninvasive method of postcontrast small intestine CT (computed tomography) examination. It combines the speed and resolution of multidetector CT scanners with enhancing properties of both ingested and intravenous contrast agent. The visualization of the intestinal wall and lumen is much better than by performing a normal CT scan or other techniques and clearly shows small intestine inflammation by displaying the thickening of the intestinal wall. The oral contrast agent is applied in several doses starting about 60 minutes prior to examination. Intravenous contrast agent is injected about one minute before examination. The doses and timings may vary, but in principle follow this basic schedule. This considerably enhances the visibility and details of the small intestine walls and already suffices for a radiologist to perform an examination and a diagnosis. Even then such an examination is very difficult even for a specialist, because the topology of the small intestine is very complicated and hard to follow. We need good preprocessing for an automatic or semiautomatic diagnosis help such as intestinal wall thickness analysis or virtual straightening of the intestine. Data quality We need thin slices because of the complicated structure and small details. But thin slices bring a lot of noise with it and are thus very hard to follow and segment. Many current segmentation algorithms need a robust edge detection. In case of an average CT enterography scan, the standard deviation of the homogeneous inner parts of the intestine (lumen filled with negative contrast agent) is bigger than the difference between the mean value of the lumen and the contrast enhanced intestinal wall. And we are talking only about the random noise present in the data and discarding the effects of acquisition artifacts. So we need a robust denoising approach to apply some proven and efficient segmentation algorithm. We will focus only on random noise, which is very apparent in the intestinal part of the body. Acquisition artifacts, such as star artifacts resulting from dense objects being present in the body are not so apparent, because unless the patient has some sort of metal implant this part of the body usually contains only the spine and upper part of pelvis and no other dense bones or objects. For example Gu et al. (2006) discusses a method of star artifact removal. 64 WDS'10 Proceedings of Contributed Papers, Part I, 64–69, 2010. ISBN 978-80-7378-139-2 © MATFYZPRESS