Projection image inpainting in X-ray CT using Relevant Neighbor Area (RNA) and Intersection of Confidence Intervals (ICI)

Abstract

Recently reported Relevant Neighbor Area (RNA) methods for projection image inpainting have achieved a drastic improvement in reducing inpainting errors simply by excluding irrelevant neighbor pixels in computing an inpainting value. This research presents how to design more accurate RNA masks in order to further improve the inpainting performance of current RNA masks. We employ the Intersection of Confidence Intervals (ICI) rule that provides spatially adaptive window size for optimal signal estimation in the presence of noise. In our case, it allows us to eliminate outlier pixels in RNA that degrades the inpainting performance, yielding an optimal RNA mask. Since the ICI rule gives an optimal neighbor area, we are free from the difficult problem of deciding the size of neighbor area for inpainting. Experimental results show that inpainting methods using more accurately determined RNA masks significantly outperform the methods that employ the current RNA masks.