Accelerated Image Reconstruction on a Cluster of Two AMD GPUs in CBCT with Non-uniform Detector Geometry

The imXgam group of CPPM has developed the XPAD3 hybrid pixel X-ray detector with 560x960 pixels (130 μm each). A Cone Beam CT scanner (CBCT) has been built, based on XPAD3 detectors, which enables volume reconstruction of a complete mouse with 65 μm resolution. However the basic unit is a chip whose width is about1cm. These are assembled into barrettes which are in turn assembled in a tiled structure in order to reduce the dead regions. This results in a geometry where the apparent position and size of the pixels are not uniform. The volume reconstruction is performed using the analytic method of Feldkamp, Davis and Kress. Reconstruction time was estimated to a few hours due to the large number of pixels and images and the added complexity of the interpolation resulting from the non-uniform geometry. The algorithm was ported to a GPU architecture based on the AMD Firestream9270 board driven via the Brook+ library. All the processing is performed in the GPU and the host CPU is used only for steering and I/O operations. For improved performance, a dual-GPU version has been developed. With this setup, the total reconstruction time is of order 2 minutes, mostly limited by disk I/O. Compared to the usual CPU implementation the gain in speed is of the order of 100