Development of an exterior-problem-based inpainting technique for metal artifact reduction in CT

Abstract

Metal artifacts have been a serious threat in the image quality and accuracy of computed tomography (CT) images both in medical and industrial applications. For decades, various strategies have been developed and refined to mitigate the effects of metal artifacts in CT. Approaches such as simple sinogram inpainting to intricate computations of iterative reconstruction methods have been explored and successful to some extent. We proposed and developed a hybrid metal artifact reduction (MAR) technique based on an exterior-problem inpainting that offers more superior images than interpolation-based inpainting and computationally more efficient than conventional iterative reconstruction algorithms. Our proposed hybrid approach is an image-based artifact subtraction which performs prior image-based sinogram inpainting, metal sinogram extraction, and metal artifact image subtraction. In short, the proposed approach was referred to as proposed IMAR. Prior image was reconstructed based on an exterior-problem concept via conventional iterative reconstruction algorithm, specifically, the total variation minimization (TVM) algorithm. Exclusively reconstructing a prior image using TVM algorithm is time-consuming. Hence, a pseudo-discrete algebraic reconstruction technique (PDART) was introduced to accelerate the reconstruction process. The missing background contribution in the metal trace was estimated after a reprojection of the prior image. Likewise, metal-corrupted sinogram can be extracted and reconstructed. Subsequently, an artifact image or a streak image was segmented and subtracted from the uncorrected image. Conventional linear interpolation inpainting MAR, or standard MAR, was also implemented for comparison. The performance of standard MAR and the proposed MAR approach with the inclusion of PDART was validated using simulation and real data studies. Image quality of MAR images was assessed by the structural similarity (SSIM) index for simulation data and standard deviation (SD) for real data studies.