Simulating X Ray Images From Deformable Shape and Intensity Models on the GPU


In medical 3D-from-2D reconstruction, the goal is to reliably reproduce three-dimensional anatomical structures from two-dimensional data like x-ray images. State of the art approaches compare a real, clinical x-ray to a high number of virtual x-ray images, which are generated from deformable 3D models. It is assumed that a deformed model represents the anatomy depicted in the clinical x-ray image, if an adequate match between its virtual x-ray and the clinical data is found. The accurate simulation of x-ray attenuation in 3D models thus plays a critical role in the 3D-from-2D reconstruction process. This work presents several fast and precise methods to generate x-ray images from statistical volumetric models, which are enriched with density information (Statistical Shape and Intensity Models, SSIMs). Previous approaches are adapted such that the x-ray attenuation is simulated entirely on graphics processing units (GPUs). Moreover, a novel method is proposed that efficiently integrates both model deformation as well as x-ray simulation into a single, GPU-accelerated operation. The presented algorithms accurately project higher order attenuation distributions, allowing flexible modeling of density within individual cells of the SSIM. Based on the OpenGL interface, a prototypical GPU implementation is provided that demonstrates the high performance of the discussed approaches.

Master thesis in Computer Science, Technische Universität Berlin