Photon-counting normalized metal artifact reduction (NMAR) in diagnostic CT
Achim Byl,Laura Klein,Stefan Sawall,Sarah Heinze,Heinz-Peter Schlemmer,Marc Kachelrieß
DOI: https://doi.org/10.1002/mp.14931
Abstract:Purpose: Metal artifacts can drastically reduce the diagnostic value of computed tomography (CT) images. Even the state-of-the-art algorithms cannot remove them completely. Photon-counting CT inherently provides spectral information, similar to dual-energy CT. Many applications, such as material decomposition, are not possible when metal artifacts are present. Our aim is to develop a prior-based metal artifact reduction specifically for photon-counting CT that can correct each bin image individually or in their combinations. Methods: Photon-counting CT sorts incoming photons into several energy bins, producing bin and threshold images containing spectral information. We use this spectral information to obtain a better prior image for the state-of-the-art metal artifact reduction algorithm FSNMAR. First, we apply a non-linear transformation to the bin images to obtain bone-emphasized images. Subsequently, we forward-project the bin images and bone-emphasized images and multiply the resulting sinograms with each other element-wise to mimic beam hardening effects. These sinograms are reconstructed and linearly combined to produce an artifact-reduced image. The coefficients of this linear combination are automatically determined by minimizing a threshold-based cost function in the image domain. After thresholding, we obtain the prior image for FSNMAR, which is applied to the individual bin images and the lowest threshold image. We test our photon-counting normalized metal artifact reduction (PCNMAR) on forensic CT data and compare it to conventional FSNMAR, where the prior is generated via linear sinogram inpainting. For numerical analysis, we compute both the standard deviation in an ROI with metal artifacts and the CNR of soft tissue and fat. Results: PCNMAR can effectively reduce metal artifacts without sacrificing the overall image quality. Compared to FSNMAR, our method produces fewer secondary artifacts and is more consistent with the measurements. Areas that contain metal, air, and soft tissue are more accurate in PCNMAR. In some cases, the standard deviation in the artifact ROI is reduced by more than 50% relative to FSNMAR, while the CNR values are similar. If extreme artifacts are present, PCNMAR is unable to outperform FSNMAR. Using either two, four, or only the highest energy bin to produce the prior image yielded comparable results. Conclusions: PCNMAR is an effective method of reducing metal artifacts in photon-counting CT. The spectral information available in photon-counting CT is highly beneficial for metal artifact reduction, especially the high-energy bin, which inherently contains fewer artifacts. While scanning with four instead of two bins does not provide a better artifact reduction, it allows for more freedom in the selection of energy thresholds.