IMPLEMENTATION OF THREE DIMENSIONAL RECONSTRUCTION AND RING ARTIFACT CORRECTION ON GRAPHICS PROCESSING UNIT (GPU)

• Main Author: (NIM : 20215021), KAMIRUL
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/22748
• Institution: Institut Teknologi Bandung

A GPU-based (Graphics Processing Unit) program has been developed to perform reconstruction process on micro CT scanner projection images. The program was implemented using FDK (Feldkamp, Davis, and Kress) which is intended to reconstruct three dimensional volume from projection images acquired using conebeam X-ray with a planar detector configuration. Based on simulation performed on GTX 780 GPU (2304 cores), the program was able to accelerate filtered-backprojection process up to 89,90 times compared to Intel <br /> <br /> <br /> CoreTM i5-4330 CPU (Central Processing Unit) implementation on 20483 voxels phantom. By varying voxels number, acceleration values were influenced by the number of reconstructed voxels and available global memory size on GPU. Maximum total acceleration value was 27,93 obtained when phantom 10243 was reconstructed on 2,05 GB GPU memory. Further result show that developed program was capable of producing acceleration values to 1.27 and 1.30 against <br /> <br /> <br /> NRecon software on medium (11203 voxels) and high (22403 voxels) resolution aquades phantom. While in low-resolution (5603 voxels), the program yields a deacceleration value of 0.85. To overcome ring artifact on the reconstructed images, the program also featured by template substraction based ring correction. From the simulation, implemented correction method was able to reduce standard deviation values on the homogeneous ROI by 22.1%, 21.5% ,23.9% for low, medium, and high resolution reconstructed phantom. Ring corrected images also evaluated by counting the number of existed ring segments using Connected Component Labeling method which was accurately applied on simple ringcontained image. Based on calculation applied to corrected images, ring segments decreased by 92.31%, 92.79%, 94.69% on corrected medium, and highresolution <br /> <br /> <br /> images respectively.