high performance computing on graphics processing units: hgpu.org

In the medical diagnosis and treatment planning, radiologists and surgeons rely heavily on the slices produced by medical imaging scanners. Unfortunately, most of these scanners can only produce two dimensional images because the machines that can produce three dimensional are very expensive. The two dimensional images from these devices are difficult to interpret because they only show cross-sectional views of the human structure. Consequently, such circumstances require highly qualified doctors to use their expertise in the interpretation of the possible location, size or shape of the abnormalities especially for large datasets of enormous amount of slices. Previously, the concept of reconstructing two dimensional images to three dimensional was introduced. However, such reconstruction model requires high performance computation, may either be time-consuming or costly. Furthermore, detecting the internal features of human anatomical structure, such as the imaging of the blood vessels, is still an open topic in the computer-aided diagnosis of disorders and pathologies. This study proposed, designed and implemented a visualization framework named SurLens with high performance computing using Compute Unified Device Architecture (CUDA), augmenting the widely proven ray casting technique in terms of superior qualities of images but with slow speed. Considering the rapid development of technology in the medical community, our framework is implemented on Microsoft .NET environment for easy interoperability with other emerging revolutionary tools. The Visualization System was evaluated with brain datasets from the department of Surgery, University of North Carolina, United States, containing 109 datasets of MRA, T1-FLASH, T2-Weighted, DTI and T1-MPRAGE. Significantly, at a reasonably cheaper cost, SurLens Visualization System achieves immediate reconstruction and obvious mappings of the internal features of the human brain, reliable enough for instantaneously locate possible blockages in the brain blood vessels without any prior segmentation of the datasets.