Metal artifact reduction using confidence maps and patch-based method
Résumé
In the field of non-destructive testing of materials, computed tomography became a good way to check defects in industrial piece production. However, tomographic analysis is difficult to achieve due to the presence of high density objects (such as metal) in most produced pieces, leading to the well-known metal artifacts in reconstructed data. In X-Ray tomography, metal artifact is characterized by a local and straight hyper-signal. This observed phenomenon is due to high attenuations of the rays in the high density materials.
Many different approaches have been proposed for metal artifact reduction (MAR) during the last decade. In the most popular case, developed methods consider the artifact reduction as a missing data problem. Currently, two kind of methods can be distinguished. The first one is based on projection completion method. Missing data, represented by the metal traces in the sinogram, are replaced by synthetic data computed by interpolation or in-painting methods. The second one considers metal artifact reduction as a penalizing term in an iterative reconstruction method. Such a method is based on local filtering assessing the coherence of a pixel with its neighborhood.
However, these methods start their process from the original projection data. In our context, only the recon- structed image is available due to clinical scanner usage. Thus, we focus our investigations on a solution based on the reconstructed image only. We propose an algorithm that reduces the artifact by directly applying a patch-based image processing. Patches preserve geometry in the image and avoid over-smoothing. However, patch-based method amends all pixels whereas only few of them have to be corrected in our case. Thus, in order to only restore corrupted pixels, we compute confidence maps from a simulated sinogram and apply them as weighting functions determining the amount of correction to perform using the patch-based method.
These maps represent hyper-signal and hypo-signal repartition generated by metal artifact in the reconstructed image. In other words, each map is a representation of the amount of hyper or hypo-signal phenomenon that occurs during the initial acquisition / reconstruction process. They are obtained using an iterative reconstruction from a simulated acquisition of metal traces. The overall output function is a combination of these two maps, where negative and positive values represent the hypo-signal and the hyper-signal, respectively.
The following of the paper is organized as follows. We first present methods selected in the literature for comparison. Then, we detail and illustrate the proposed method. Finally, we compare and discuss reconstruction results obtained by our method and methods selected in the literature.